Innovations for Community Services pdf pdf


1 year ago
Full text

  Michal Hodoň Gerald Eichler Christian Erfurth Günter Fahrnberger (Eds.)

Communications in Computer and Information Science 863

Innovations for Community Services 18th International Conference, I4CS 2018 Žilina, Slovakia, June 18–20, 2018 Proceedings


in Computer and Information Science 863

  Commenced Publication in 2007 Founding and Former Series Editors: Alfredo Cuzzocrea, Xiaoyong Du, Orhun Kara, Ting Liu, Dominik Ślęzak, and Xiaokang Yang

  Editorial Board

  Simone Diniz Junqueira Barbosa Pontifical Catholic University of Rio de Janeiro (PUC-Rio), Rio de Janeiro, Brazil

  Phoebe Chen La Trobe University, Melbourne, Australia

  Joaquim Filipe Polytechnic Institute of Set úbal, Setúbal, Portugal

  Igor Kotenko St. Petersburg Institute for Informatics and Automation of the Russian Academy of Sciences, St. Petersburg, Russia

  Krishna M. Sivalingam Indian Institute of Technology Madras, Chennai, India

  Takashi Washio Osaka University, Osaka, Japan

  Junsong Yuan University at Buffalo, The State University of New York, Buffalo, USA

  Lizhu Zhou Tsinghua University, Beijing, China More information about this series at

  • Michal Hodoň Gerald Eichler Christian Erfurth Günter Fahrnberger (Eds.)

  Innovations for Community Services

18th International Conference, I4CS 2018

Žilina, Slovakia, June 18

  • –20, 2018 Proceedings
Editors Michal Hodoň Christian Erfurth University of Žilina EAH Jena Žilina Jena Slovakia Germany Gerald Eichler Günter Telekom Innovation Laboratories University of Hagen Deutsche Telekom AG Hagen Darmstadt, Hessen Germany Germany

ISSN 1865-0929

  ISSN 1865-0937 (electronic) Communications in Computer and Information Science

ISBN 978-3-319-93407-5

  ISBN 978-3-319-93408-2 (eBook) Library of Congress Control Number: Applied for © Springer International Publishing AG, part of Springer Nature 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the

material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,

broadcasting, reproduction on microfilms or in any other physical way, and transmission or information

storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now

known or hereafter developed.

The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication

does not imply, even in the absence of a specific statement, that such names are exempt from the relevant

protective laws and regulations and therefore free for general use.

The publisher, the authors and the editors are safe to assume that the advice and information in this book are

believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors

give a warranty, express or implied, with respect to the material contained herein or for any errors or

omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in

published maps and institutional affiliations. Printed on acid-free paper This Springer imprint is published by the registered company Springer International Publishing AG part of Springer Nature.



  The International Conference on Innovations for Community Services (I4CS) had its 18th edition 2018. It had emerged as the Workshop on Innovative Internet Community Systems (I2CS) in 2001, founded by Herwig Unger and Thomas Böhme, and continued its success story under its revised name I4CS in 2014. We are proud to have reached again the original number of scientific presentations, combined with a great social conference program.

  The selection of conference locations reflects the conference concept: Our members of the Technical Program Committee (TPC) can offer suitable locations. In 2018, the Steering Committee had the honor of handing the organization responsibility over to Michal Hodoň and, therefore, of determining a Slovakian venue for the first time in the history of the conference. The University of Žilina was a remarkable place for offering a perfect climate to make the motto “Relaxation Teams Communities” happen.

  I2CS published its first proceedings in Springer series Lecture Notes in Computer Science series (LNCS) until 2005, followed by the Gesellschaft für Informatik (GI), and Verein Deutscher Ingenieure (VDI). I4CS commenced with the Institute of Elec- trical and Electronics Engineers (IEEE) before switching back to Springer’s Commu- nications in Computer and Information Science (CCIS) in 2016. With 1,473 chapter downloads from SpingerLink for CCIS Vol. 717, publishing the I4CS proceedings of 2017, we envisaged an increasing result. I4CS has maintained its reputation as a high-class C-conference at the CORE conference portal

  ranks/?search=I4CS&by=all .

  The proceedings of I4CS 2018 comprise five parts that cover the selection of 14 full and three short papers out of 38 submissions. Interdisciplinary thinking is a key success factor for any community. Hence, the proceedings of I4CS 2018 span a range of topics, bundled into three areas: “Technology,” “Applications,” and “Socialization.”

  Technology: Distributed Architectures and Frameworks

  • Data architectures and models for community services
  • Innovation management and management of community systems
  • Community self-organization in ad-hoc environments
  • Search, information retrieval, and distributed ontologies
  • Common data models and big data analytics

  Applications: Communities on the Move

  • Social networks and open collaboration
  • User-generated content for business and social life
  • Recommender solutions and context awareness
  • Augmented reality and location-based activities
  • Intelligent transportation systems and logistic services
VI Foreword

  Socialization: Ambient Work and Living

  • eHealth challenges and ambient-assisted living
  • Intelligent transport systems and connected vehicles
  • Smart energy and home control
  • Digitalization and cyber-physical systems
  • Security, identity, and privacy protection

  Many thanks to the 19 members of the TPC representing 12 countries for their valuable reviews, especially the chair, Christian Erfurth and, secondly, to the publi- cation chair, Günter Fahrnberger, who fostered a fruitful cooperation with Springer.

  The 19th I4CS will be organized by the Ostfalia University of Applied Sciences and will take place in Wolfsburg/Germany in June 2019. Please check the permanent conference URL for more details. Applications of prospective TPC members and potential conference hosts are welcome at

  April 2018 Gerald Eichler



  Žilina is the natural center of northwestern Slovakia, which ranks among the largest and most important cities in Slovakia. It is located in the valley of the Váh River, surrounded by the beautiful mountain ranges of Malá Fatra, Strážovské vrchy, Súovské vrchy, Javorníky, and Kysucká vrchovina. The National Park of Malá Fatra comprises famous gorges, rock peaks, and an attractive ridge tour. The main subject of protection is the territory with varied geological history and dissected relief forms, rare and precious biocenoses, flora and fauna, and the exceptional value of the forest and mountain compounds with precious dwarf pinewoods, and rapacious animals, such as the wolf, lynx, or bear.

  Žilina is a center of significant political, cultural, sport, and public health-care institutions. Its economic potential can be proven by the fact that Žilina has the second highest number of traders per thousand inhabitants. As for the number of joint stock companies and limited companies, Žilina holds third position in Slovakia. Nowadays, the city of Žilina represents a dynamic development accelerated by KIA Motors Slo- vakia investments. However, the city is not only a center of car production, but together with the Upper Váh River Region (Horné Považie) it is an interesting tourist desti- nation. The city of Žilina is a center of theaters, museums, galleries, parks, and sports facilities. Its historical center is crossed by one of the longest and the most beautiful pedestrian zones in Slovakia.

  The University of Žilina was founded in 1953 by separating from the Czech Technical University in Prague, followed by its renaming to the University of Trans- port and Communications. Later in 1996, after broadening its fields of interest and other organizational changes, it was renamed as the University of Žilina. In its over 60 years of successful existence, it has become the alma mater for more than 70,000 graduates, highly skilled professionals mostly specializing in transport and technical fields as well as in management, marketing, or humanities. The quality and readiness of the graduates for the needs of practice is proved by long-term high interest in hiring them by employers that cooperate with the university in the recruitment process.

  A stopover in the Malá Fatra Mountains offers unforgettable experiences enhanced through the selected venue of the Village Resort Hanuliak as a unique wellness resort located in the beautiful environment of the Malá Fatra National Park. The picturesque village of Belá is located only 20 km away from the city of Žilina.

  We hope that all attendees enjoy the fruitful, friendly, and relaxed atmosphere during the conference. We trust they will gather professional experiences and be happy to come back in the future. April 2018

  Michal Hodoň Organization Program Committee

  Marwane Ayaida University of Reims Champagne-Ardenne, France Gilbert Babin HEC Montréal, Canada Gerald Eichler Deutsche Telekom AG, Germany Christian Erfurth Jena University of Applied Sciences, Germany Günter Fahrnberger University of Hagen, Germany Hacène Fouchal University of Reims Champagne-Ardenne, France Sapna Gopinathan Coimbatore Institute of Technology, India Michal Hodoň University of Žilina, Slovakia Peter Kropf University of Neuchâtel, Switzerland Ulrike Lechner Bundeswehr University Munich, Germany Karl-Heinz Lüke Ostfalia University of Applied Sciences, Germany Phayung Meesad

  King Mongkut’s University of Technology North Bangkok, Thailand

  Raja Natarajan Tata Institute of Fundamental Research, India Frank Phillipson TNO, The Netherlands Srinivan Ramaswamy ABB, USA Joerg Roth Nuremberg Institute of Technology, Germany Maleerat Sodanil

  King Mongkut’s University of Technology North Bangkok, Thailand

  Leendert W. M. Wienhofen

  City of Trondheim, Norway Ouadoudi Zytoune Ibn Tofail University, Morocco



   . . .

   äthe . . .

   J érémie Bosom, Anna Scius-Bertrand, Haï Tran, and Marc Bui . . .

   Geoffrey Wilhelm, Hac ène Fouchal, Kevin Thomas, and Marwane Ayaida

  . . .

  . . .

   Michal Kvet and Karol Matiasko . . .

   Emilien Bourdy, Kandaraj Piamrat, Michel Herbin, and Hac ène Fouchal . . .

  . . .

   Veronika Olešnan íková, Ondrej Karpiš, Lukáš Čechovič, and Judith Molka-Danielsen

  Julian Knoll, Rainer Gro ß, Axel Schwanke, Bernhard Rinn, and Martin Schreyer

  Andreas Lommatzsch XII Contents


  ès Dinant, Thomas Vilarinho, Ilias O. Pappas, Simone Mora, In Jacqueline Floch, Manuel Oliveira, and Letizia Jaccheri


  Karl-Heinz L üke, Johannes Walther, and Daniel Wäldchen . . . Marcus Wolf, Arlett Semm, and Christian Erfurth

   Sven Von Hollen and Benjamin Reeh


  Janusz Furtak, Zbigniew Zieliński, and Jan Chudzikiewicz

  Jakub Hrabovsky, Pavel Segec, Marek Moravcik, and Jozef Papan

  R óbert Žalman, Michal Chovanec, Martin Revák, and Ján Kapitulík

  Architectures and Management


Microservice Architecture Within

In-House Infrastructures for Enterprise

Integration and Measurement:

An Experience Report

  ( ) B

  Sebastian Apel , Florian Hertrampf, and Steffen Sp¨ athe


Friedrich Schiller University Jena, 07743 Jena, Germany



  The project WINNER aims to integrate and coordinate

electromobility used through carsharing, the energy consumption of

tenant households and the local production of electricity, e.g., by

integrating photovoltaic systems into a smart local energy grid. While

the various components correspond to the currently available standards,

the integration has to be realised via a data processing and storage

platform, the WINNER DataLab. The goal of this platform is to provide

forecasts and optimisation plans to operate the WINNER setup as

efficiently as possible. Each data processing component is encapsulated

as a single service. We decided to use a microservice architecture

and further an execution environment like container instances within

a cloud infrastructure. This paper outlines the realisation as well

as a report of our experiences while realising this project related

microservice architecture. These experiences focus on development

complexity, modifiability, testability, maintainability and scalability as

well as dependencies and related lessons learned. Finally, we state,

that the practical application of setups like this helps to concentrate

on business models. It supports decoupling, helps in development to

focus on the essential things and increases efficiency in operation, not

least through good opportunities for scaling. However, it is required

to mastering the complexity which currently requires clean planning,

experience and a coordinated development process.

  Keywords: Smart grid Internet of things Microservice architecture

· ·

  Experience report Measurement infrastructure ·

1 Introduction

  Imagine a modern urban area with tenant households. Photovoltaic systems produce electricity; each flat knows when electricity is available. So, scheduling of consumers is possible as well as the management of electric cars that are charged when electricity is available. In contrast, electricity is offered when the

4 S. Apel et al.

  car will not be used soon. The implementation of such smart grids, especially its network of actors and sensors, which is known as the internet of things (IoT), is a complex task

   ]. IoT is stated as the next big step in internet technology



], and there is a large number of different and heterogeneous devices to handle



]. One way to address this integration and measurement task are microservice

architectures and cloud computing infrastructures.

  Our research project “Wohnungswirtschaftlich integrierte netzneutrale Elektromobilitat in Quartier und Region” WINNER

   ] aims to integrate and

  coordinate electromobility used through carsharing, the energy consumption of tenant households and the local production of electricity, e.g., by integrating photovoltaic systems into a smart local energy grid. Our primary goal of this project is avoiding the injection of electrical power into higher grid levels, which means Level 7 for local distribution and above referring to

   ]. While

  the resulting installation uses currently available components, our focus within this paper is on creating an integration and measurement platform to gather, analyse and provide information from the installation. The objective of this platform is to provide forecasts and optimisation plans to operate the test setup as efficiently as possible. Due to the various endpoints and our agile process of implementing the overall system, we want to focus on infrastructure, architecture and development aspects to realise systems like this.

  The so-called WINNER DataLab (WDL) is the integration and measurement platform of all project related sources and sinks, e.g., devices within the installation as well as external services. As evaluated within

  , the

  architectural backbone technology we want to use is Apache Camel

   ]. This

  backbone allows to integrate various systems and to coordinate the resulting data flows. Further, each data processing component is encapsulated as a single small service and wired together by using representational state transfer (REST) and messaging services. This setup implies the use of a microservice architecture and further some kind of execution environment, e.g., isolated and independent container instances within a cloud infrastructure. But, there is a project related requirement, which states that the whole setup has to run on in-house infrastructure. This requirement is motivated by security issues to keep sensitive data within quarters near tenant households. So, the realised setup has to be deployable on locally executed hardware and data should not leave the area.

  The following paper outlines the efforts in planning, development, deployment and operation of the WDL. Based on the necessity of in-house operation, this concerns, in particular, the components for a microservice architecture as well as the services for the operation of a compact cloud computing infrastructure and tools to support development and deployment. This applies to small sized setup and does not relate to large infrastructures, provided by, e.g., Amazon, Google or Microsoft. This experience report focus is on required components to realise a fully working setup. Thus, the evaluation outlines our experiences regarding development complexity, modifiability, testability, maintainability and scalability as well as dependencies and related lessons learned.

  Microservice Architecture Within In-House Infrastructures


  The remaining paper is organised as follows. In Sect.

   we will go into

  details about related work on microservices, measurement infrastructure and architectures for the IoT. Section

   deals with details of the project, how the

  different components interact as well as our required components to manage measurement and data flow. The following Sect.

   goes into details about required components to manage a microservice infrastructure and how they interact.

  Finally, Sect.

   discusses our experiences regarding the already mentioned areas, like complexity, modifiability and scalability.

2 Related Work

  Smart Cities, IoT and Cloud Computing, are ongoing research and industry efforts. These efforts are aiming at development as well as deployment standards and best practices for designing systems and platforms. More generally, IoT architectures are discussed in

   ]. They point out that a flexible layered

  architecture is necessary to connect many devices. As an example, classical three- and five-layer architectures are mentioned here, e.g., as used within

   ], as

  well as middleware or SOA-based architectures. Further,

   ] presents another

  overview of the activities done in Europe towards the definition of a framework “for different applications and eventually enable reuse of the existing work across the domains”. Besides this discussion about architectures,

   ] demonstrates the

  integration of IoT devices within a service platform which uses the microservice architecture for this approach, which can be understood as a specific approach for service-oriented architecture (SOA)

   , P. 9]. However, thinking about

  microservices requires regarding principles and tenets

  , like fine-grained interfaces, business-driven, cloud-native, lightweight and decentralised.

  In addition to IoT, measurement systems for IoT can also be considered. They also have to integrate different end systems. Further, they have to record different measured values and provide interfaces for analyses and calculations.

  For this, approaches like SOA or event-driven architecture (EDA) can be taken up, as demonstrated in

   ]. This approach uses SOA and EDA in combination

  with an enterprise service bus (ESB). Using the microservice architecture can be seen in

   ] as loosely coupled components by using the enterprise

  application integration (EAI)

  P. 3]. They describe a reference architecture

  using microservices for measurement systems, which connects required data adapters as well as calculation and storage services, one more time through an ESB.

  In summary, the shown references deal in particular with the design of IoT and measurement infrastructures. They use SOA, EDA or microservices, combine them with an ESB for the decoupled exchange of events or connect the various components directly to each other. The approaches appear domain-specific and list the necessary components; less attention is paid to the operation, the practices and development of the platforms.

  6 S. Apel et al. External Power Grid

Photovoltaic Car Space Management

Meter Inverter Charging Infra.

Meter Clamp Clamp Meter

  Mobility Producer

Cluster Controller Electric Vehicle

Infrastructure Signal Converter Meter Clamp

Residents Device Connectivity

Bus Quarter Meter Clamp Field Energy Management WINNER DataLab Connectivity Management

Other Cables Carsharing Controlling

Data Energy Weather Historical Weather Forecast Data Stream Processing Carsharing Bookings

  Analog Legend Ethernet Services Energy Charge Forecast Fig. 1.

  Overview of all components of the demonstrator in the WINNER project.

  3 WINNER Setup

  WINNER aims to integrate and coordinate electromobility used through carsharing, the energy consumption of tenant households and the local production of electricity, e.g., by integrating photovoltaic systems into a smart local energy grid. Thus, there are different components, as they are currently available, for the acquisition of measured values. Figure

   shows all related

  components within this setup. This architectural overview is derived and discussed in

  . These components are divided into six parts. The first part

  is related to the photovoltaic systems. It contains several photovoltaic panels, their inverters for connecting to the power grid and a controller for managing the system. Also, measuring points are provided for recording the amount of electricity generated, in particular, a meter and a clamp. The second part is related to the tenant households. These also include meters and clamps for each household, as well as a meter and clamp for installations used by all residents. The third part is about mobility. In addition to the meter and clamp, the charging infrastructure and car space management are required here. The fourth part refers to connectivity. This applies in particular to the acquisition of measured values from meters and clamps and the providing of data for processing components. The fifth part is related to external services. The project took into account weather services, car sharing services and electricity exchange services,

  Carsharing Bookings Device Connectivity Microservice Architecture Within In-House Infrastructures

<< use >> << use >>

Data Accessor Services

<< use >> Visualization

  7 Carsharing Controlling Realtime Measurement Values Weather Historical Weather Forecast Optimization Roadmap Forecasts Message Device Connectivity Master Data Runtime Data

Energy Charge Forecast Subsystem Carsharing Controlling

Energy Management Processing Energy Management

Input Adapters Services Output

Fig. 2.

  Generalized representation of the involved adapters and services for the WDL.

  e.g., provided by European Energy Exchange (EEX). Finally, all five parts have to be integrated for further analyses and calculations within the already mentioned WDL as the sixth part of this setup. This sixth part also contains the component Energy Management, which uses WDL to access the data flows and databases via events and interfaces to carry out further analyses, such as the preparation of forecasts and optimisation plans.

  Based on the components in the demonstrator mentioned above, an architecture for gathering, processing and analysis of various data streams can be designed. This data stream processing platform for enterprise application integration, the WDL, is visualised on its Level 1 component view in Fig.


  illustration on the left-hand side shows the expected inputs of the demonstrator and the external services, which provide in particular the various measured values from installations such as meters and clamps, as well as information on carsharing, weather and electricity exchange. These adapters in the WDL generate events on different data streams and are made available to other services via a message service. The right-hand side of the figure is dominated by advanced services that gather and process events from the various message queues. This gathering applies, for example, to services that continuously persist events in databases for time series or master data, as well as services that generate events for the demonstrator by continuously evaluating the data streams. Also, there are services for accessing persistent data in the databases that are not specified in detail. These can be used for mapping, enrichment and evaluations, for example. The structure is comparable to the reference architecture found in .

4 Microservice Infrastructure Setup

  While up to this point the microservice architecture and the components are comparable to the publication mentioned in Sect.

  the question arised how such a setup can be realized, orchestrated and operated in an in-house infrastructure.

8 S. Apel et al.

  In-house, in this case, means an underlying Infrastructure as a Service (IaaS), which offers the availability of hardware and associated software. This IaaS covers server, storage and network, as well as operating systems virtualisation technology and file systems

   ]. There are some virtual machines. The necessity

  for in-house operation lies in the continuous but secure processing of sensitive tenant household data to make the necessary decisions for the quarters.

  Within our use case, we decided to use a cluster of docker engines. In this cluster, the components required for our microservice architecture are operated in individual containers. This individualisation means that a container always corresponds to precisely one service, services to integrate various components as well as analyse data and provide plans for optimised usage of energy. While this is state-of-the-art, the question of how to bring up this bunch of containers into execution remains, especially when thinking about tenets and principles.

  First of all, we need a management and orchestration toolset. With the increasing number of microservices, the effort for administration increases. We used docker compose

   ] to orchestrate multiple containers and configure the

  basic setup. Further, we have used a web tool called Portainer

  to monitor

  and operate our containers. The primary purpose of this application are a health check and possibilities to stop or restart services. The service and the communication between the service and the Docker engines must be adequately secured and protected against external access. In this case, we have set up a communication layer based on internal network communication as well as transport layer security (TLS). One instance within the cluster is sufficient to manage it.

  The next topic is related to management of container related events, e.g., log messages. For this purpose, we have used an ELK Stack

  . Using logstash, the

  various messages of the containers are recorded, forwarded to the Elasticsearch database and finally viewed within Kibana. Logstash is recommended for each cluster node, especially when processing file-based logs. The Elasticsearch as database and Kibana for viewing is only necessary once. With increasing event traffic, the database may be scaled. Additionlly, collection metrics about CPU, RAM and Storage usage as well as incoming requests and method invocations is advisable. For this application case, we used Stagemonitor

   ] for Java-based

  services, which is executed within the Java VM of each services, collects the desired information and pushes them also to the Elasticsearch database.

  A service discovery component realises connecting services or containers. Eureka from Spring Cloud Netflix

  is used for this purpose. This service

  provides a directory and makes it possible that each service can register, services can find other services, health check for services and various services do not have to be dependent on specific configurations. One instance per cluster is minimum; multiple can be used mainly to separate environments, e.g. staging and production systems. For services that want to use other services, a request to the discovery service is necessary, as well as the decision which one should be requested from the set of available services. The task can be realised with the help of a client-side load balancer such as Ribbon in Spring applications

  or Microservice Architecture Within In-House Infrastructures



  in case of Node.js. An alternative would be to rely on the service

  discovery strategies of the execution environment. For example, Kubernetes


  provides a grouping of services based on a label selector to create a virtual and addressable endpoint. This endpoint has its own IP and can be used by services that want to talk to a service from the group. For the requesting service, it is not clear which service from the group will ultimately process the request.

  In addition to the service discovery, a gateway is required to provide service interfaces for external and frontend clients. In our case, we use Zuul

  as a

  gateway and thus offer external access to HTTP-based service interfaces. This gateway uses the Service Discovery component to coordinate communication between clients and microservice instances. One instance per cluster is minimum.

  For the configuration of the individual services, it is necessary to schedule a central configuration service. The task of the service is to return a set of key-value pairs, broken down by application, profile and label, which reflect the configuration of the service. Due to the use of Spring Cloud Components like Eureka and Zuul, the tool Cloud Configuration

  from Spring was used in

  our application from the same toolbox. Only one instance in overall is required because the service allows differing between application profiles and labels.

  Services that publish user-specific interfaces require securing them. The use of OAuth2 and the use of access tokens, as well as their verification against the OAuth2 service by the respective services, are suitable for this purpose. Alternatively, the use of JWT is also possible. The combination of JWT and OAuth2 is possible and avoids the communication between service and OAuth2 server to check the tokens.

  Further tools are recommended to support the service development. This recommendation applies, for example, to a source code management environment. It serves for versioning, coordination of changes and control of workflows within the development team. In our case, a Gitlab is used for this purpose, which offers repositories, simple management for documents, issue tracking, as well as the integration of additional tools for communication and process support.

  Furthermore, the usage of a continuous integration platform makes sense. The aim is the continuous merging of new developments. Events in the repositories are used for this purpose, for example, and lead to construction and subsequent test processes to identify possible errors at an early stage after integration. In our tested setup, we chose the continuous integration platform integrated into Gitlab and deployed a pool of Gitlab-CI-Runner, which schedules the build and test jobs created by events from repositories.

  The topic of continuous delivery is directly linked to the build and test phase. If the preceding process is successful, executable containers are created and published in a private registry. However, until now, we do not make use of the following step of continuous deployment.


   visualises the outlined components used within initial sequences for

  discover and configuration (1) as well as an API call for “SomeService”. The first sequence covers registration of this new service with the discovery service as

10 S. Apel et al.

  Load Register and Heartbeat 1 Token Check 2.2.2 1.2 Client-side Load Balancing

Discovery Client

2.2.1 1.1 Discovery Server Eureka OAuth2 Server Load and Change SomeServiceFunction Model 2.2.3 OAuth2 Client Discover OAuth2 and Message Queue Instance 2.1 Discover SomeService Instance


Database Server Broadcast 2.2.4 Call SomeService 2.2 Client-side Load Balancing 2 Gateway Server Change State


Call SomeService Zuul Message Queue with Token Fig. 3.

  Example on communication flow in case of start up and calling a service API.

  well as the initial discovery of the configuration service and the further loading of related configuration. The second sequence covers an exemplary API call. This one starts with a request at the gateway. The gateway must then discover the actual service instance and forward the received request. The “SomeService” instance itself has to discover an endpoint to validate the token (especially if you are not using something like JWT) as well as other required services, a message service in this example. The instance finally validates the token and handles its business logic “SomeServiceFunction” which might result in a database call. Finally, the instance broadcasts some information through the already discovered message queue. The resulting response to this request will be transported back to the client.

  Monitoring Staging Productive

Container Engine Container Engine Container Engine

  Log Event Service Service Service Service 1 n ... ... 1 n Database Log Event Service Service Service Service Discovery Gateway Discovery Gateway

  Visualizer Container Messaging Messaging Log Log Management Service Service

  Processor Processor Authorization

Source Issue Continious Cloud


  Support Repository Tracking Integration ation


Fig. 4. Visualization of the in-house infrastructure for operating the services.

  Microservice Architecture Within In-House Infrastructures



   combines the outlined components within a deployment plan as

  we would do it. We have currently distributed these containers and our encapsulated microservices manually across the already mentioned and available IaaS infrastructure. This setup contains a staging and a production environment, especially to test new builds and to prevent side effects, e.g., register untested service names within the namespace of the system used in productive. This deployment uses some components only once because they are not required to replicate.

  5 Evaluation of Experiences Our microservice architecture is partly realised as described in Sect. Within this

  evaluation, we want to focus on our experiences and lessons we have learned when realising an architecture setup like this. This outline covers development complexity, modifiability, testability, maintainability and scalability as well as dependencies, development skills and related learned lessons. An overview of the experiences are listed in Table and are presented in detail below.

  The realisation currently contains 17 service containers. There are six integration services, e.g., for interfaces like carsharing, smart meters, weather and energy prices as well as eleven infrastructure services, e.g., a timeseries database, a NoSQL database, messaging, discovery, gateway, cloud configuration and an Elasticsearch stack with Kibana for logging. More integration and analysis related services will be added in the future.

  The first outlines are related to development complexity. Developing and deploying integration services as they are mentioned above help to focus on small and fine-grained tasks. So, the complexity per service task is lowered. Thus, the resulting services are quite easy to handle. However, the application of the whole microservice infrastructure of this “integration services”, by achieving “isolated state, distribution, elasticity, automated management and loose coupling”


  requires additional concerns. These concerns and their complexities are mainly related to configuration, discovery, gateway and logging services. The developer should be trained to gain knowledge about tools and best practices of loosely coupled services. Otherwise, at least a provided tool stack which encapsulates these topics has to be used. Remarkable, these topics have to be considered within monoliths as well. However, they are getting much more impact within a highly decoupled and distributed microservice architecture. The complexity increases in case of small teams. While in theory teams specialise in the services they are given responsibility for, small teams work on the entire service composition. Therefore, all services must be developed and maintained. As we have observed, this may cause developers to get lost in the many fine-grained service implementations. This should be taken into account and, if necessary, compensated by clear processes and issue management.

  Modifiability is our second topic to review. If services are modelled around business concepts, publish well formed API and development processes are highly

12 S. Apel et al.


Table 1. Comparison of advantages and disadvantages of experience.

Advantages Disadvantages Complexity Focus on small and fine-grained tasks Isolating may requires additional dependencies Complexity per service task is lowered Few Teams for all services within a distributed architecture may cause higher complexity Modifiability

  Service focus on bounded context Infrastructure and backbone decisions may have large impacts Features change independently Testability Unit, Acceptance and Integration Tests are application-specific and clean Management of the isolated and separated networks Frequently switching of execution environment Maintainability

  Significantly simplified for developers and operators, especially in case of clean DevOps Requires extended tools to monitor and visualize communication flows for bug tracking

  Scalability Replicas can be efficiently initialised Caching strategies may require additional handling Works out of the box Partitioning of data streams, their configuration and distribution of analysation tasks among instances Handle analytical state distribution to new instances in case of failures

  Dependencies Each service can maintain its own dependency set Additional infrastructure components are required Dependency quantity can quickly become large automated, changes within services seem to be trivial and mostly easy to do.

  Thus, the modifiability is quite nice in the business use case. Apart from that, the developer should be careful while doing infrastructure and backbone decisions as well as modifications. Changing critical components which, e.g., helps to register and discover services, leads to modifications of all services. So, the system setup has to be carefully considered.

  Microservice Architecture Within In-House Infrastructures


  Testability is the third topic for outlines of experiences. The amount of use cases and tasks a microservice is used for should be well-defined and delimited. Thus, testing can be done cleanly regarding white and black box tests, e.g., testing the model itself as well as the published APIs, primarily if they are described within unit tests. Additionally, we noticed challenges and understanding issues while working with microservices regarding communication networks, visibility and API accessibility. For example, the developer is using a dedicated environment while developing, boots required infrastructure services as documented and starts working on its current feature. As suggested, each runtime test should be done by compiling the source, building the container and deploy this one to the test environment. This building takes time, much more time than simply press run within the integrated development environment (IDE). As we noticed, some developers like to preserve this possibility to merely run and test applications, and move them into containerised environments after they know it is running. We do not want to weight which development style is better; we want to retain that developers should regard the isolated and separated networks, and that they have to manage changes which have to be done to connect different execution environments.

  The next outlines are related to maintainability. Maintenance of individual services is significantly simplified. Administrators can easily monitor the services through a wide variety of tools and event management systems; for example, errors can be tracked for individual services. Other tools may be required to look at errors in a transaction or processing chain.

  The scalability, our next topic, can be realised well, particularly by horizontal scaling. Developed services that are available as containers can be replicated. The resulting replicas can be efficiently initialised if they can obtain the necessary configuration through a service. Thus, merely starting and anything else should work out of the box. However, if scaling is used and the services work with caching strategies, precautions must be considered. For example, as long as a service instance is available and not overloaded, it may be useful to ensure that requests of the same context are processed by the same service instance when distributing the load. For example, requests from the same user and related queries for third-party services could be cached locally more efficiently. Alternatively, if caching infrastructure is used, e.g., via a Redis database, it is necessary to consider how to clean up the cache. Besides, especially in services for data stream analysis, challenges arise in scaling. These scaling challenges happen, for example, when partitioning of the data stream takes place. In this case, for example, individual service instances could take over a subset of the partitions. You have to think of the method, the volume of partitions to be considered is communicated to a single service instance, especially if instances of the same service use the same configuration. Further challenges arise when time windows are considered. In this case, it has to be considered whether gaps in the analysis are justifiable. If they are not, the intermediate results for the observation interval should be persistent, so that other instances could use them as a basis. Finally, it is necessary to consider message services during scaling.

  14 S. Apel et al.

  For example, in the case of topics, it is sometimes desirable for a service to be informed to derive actions. If this service is scaled, new messages in the topic will have two instances acting in the same way. However, if other services are listening on the topic, it is not possible to simply switch to a queue. For this purpose, service groups are necessary within the message service. The distribution of a message only once to a service of a group is realised in that way using tools like Apache Kafka ], which directly offers this option.

  Experiences with dependencies are our next topic to review. Microservices have to be “decentralized” and “independently deployable”

  . To achieve this,

  additional tools are necessary as shown. However, their use in a specific service usually requires dependencies to apply the tools. The resulting dependency quantity can quickly become large. For example, if the demonstrated stack is used in a node.js service, it requires up to 10 different libraries for configuration, discovery (register and search), database access, HTTP-based API and security. Project templates or meta toolkits are conceivable but may be necessary for each service development tool.

  Also, four experiences have arisen which cannot be directly assigned to the previous categories. The first one focuses on “discover config service” or “config discovery service”? Central and essential infrastructure components are created which must be assigned to a service. In our current preferred setup, this is the Discovery service, which can be used to find a configuration service. However, it would also be conceivable the other way around. The second experience is about schema management. Schema management and its inclusion to derive client and server stubs should be considered. In our experience, services should provide schemas, but not generated stubs. This is due to the free choice of tools and programming languages for the realisation of microservices, as well as the self-responsibility of the same services. The third experience is related to hardware requirements. Most of our currently developed services are realised in Java and use Apache Camel. However, because of the specific runtime characteristics of Java, wrongly or in unbalanced configurations of heap space usage may cause a huge amount of unused resources, e.g. RAM. Analyses and evaluations of the load and idle phases during runtime can be helpful for fine-tuning configuration. Compared to monoliths, however, a finer configuration is possible and the resource usage can be directly adapted to the service. The forth and last one is about gateway configuration. While “’a culture of automation’ in testing and deployment”

   ] should be achieved, there are

  components to be careful about. For example, the operation of a gateway like Zuul for offering external clients the possibility to access microservices APIs should be carefully configured not to expose sensitive functionalities.

  6 Discussion

  The setup and experiences regarding development complexity, modifiability, testability, maintainability and scalability as well as dependencies and related lessons learned refer to the outlined use case within the project WINNER.

  Microservice Architecture Within In-House Infrastructures


  While they cannot be generalised directly, they outline challenges that are partly already known by monolithic architectures (like discovering services or logging management) or by specific microservice architectures (like client-side load balancing or cloud configuration). It is particularly noticeable to us that the freedom of choosing development environments and the independence of services in contrast to the necessary components within the service composition to achieve this independence will create library related dependencies or development efforts which have to be taken into account. The creation of microservices within a predefined environment is convenient for developers as long as the contact with the infrastructure is low and the amount of services involved in developing new features remains manageable. But, there are many dependency decisions which have to be made carefully. They require to think about how to avoid issues in later project states. Our perception in this use case is that more independence for core functionalities, like discovery and cloud configuration, is necessary. Spring-Cloud and Spring-Boot provide a nice way to do so and hide micro service related infrastructure tasks. However, the developer has to know about that and must use these tools carefully to make the proper decisions. Further, if this comfort environment is left, it can quickly lead to extensive adaptations to meet specific infrastructure requirements. Our experiences suggest that components in the field of microservices require more standardisation. The influence of the decision for central infrastructure services should be reduced. Toolkits can help in the short term, but in the long term development tools are conceivable that completely cover the various topics. Nevertheless, our application cases also show clear strengths for use. This concerns the complexity reduction of a business case within a service, as well as the good possibilities for scaling and testing.

7 Conclusion

  The objective of this paper was to describe a measurement infrastructure within the context of the research project WINNER, to realise it by using a microservices architecture and report experiences from this process. WINNER aims to integrate and coordinate electromobility used through carsharing, the energy consumption of tenant households and the local production of electricity, e.g., by integrating photovoltaic systems into a smart local energy grid. The necessary platform for data processing, our WDL, has to consider various integration, coordination and analysis tasks. While classified as non-hard real-time measurement infrastructure for data stream analysations, we want to realise this use case within a microservice architecture motivated by


] which loosely couples services through a message bus. Implementing a

  system of this type, however, requires additional infrastructure components to meet the typical characteristics of microservices, such as a high degree of automation, decentralisation, independent deployment and good monitorability


]. Mastering this complexity currently requires clean planning, experience and

  a coordinated development process. The various challenges about development

  16 S. Apel et al.


9. Al-Fuqaha, A., Guizani, M., Mohammadi, M., Aledhari, M., Ayyash, M.: Internet

of things: a survey on enabling technologies, protocols, and applications. IEEE

Commun. Surv. Tutorials 17(4), 2347–2376 (2015) Fourthquarter.


17. Krˇco, S., Pokri´c, B., Carrez, F.: Designing IoT architecture(s): a European

perspective. In: 2014 IEEE World Forum on Internet of Things (WF-IoT), pp.


16. Ibsen, C., Anstey, J.: Camel in Action, 2nd edn. Manning Publications Co.,

Greenwich (2018)


15. Hertrampf, F., Apel, S., Sp¨ athe, S.: Requirements and architecture concept for a

data processing system in local area smart grid. J. Electr. Eng. 6 (2018)


14. Choudhary, G., Jain, A.K.: Internet of things: A survey on architecture,

technologies, protocols and challenges. In: 2016 International Conference on Recent Advances and Innovations in Engineering (ICRAIE), pp. 1–8, December 2016.


13. Chemnitzer Siedlungsgemeinschaft eG: WINNER-Projekt (2017). 01 February 2018


12. Bhardwaj, S., Jain, L., Jain, S.: Cloud computing: a study of infrastructure as a

service (IAAS). Int. J. Eng. Inf. Technol. 2(1), 60–63 (2010)


11. Apel, S., Hertrampf, F., Sp¨ athe, S.: Evaluation of architectural backbone

technologies for winner datalab. In: Proceedings of the Sixth International Conference on Smart Cities, Systems, Devices and Technologies, Venice, pp. 35–43, June 2017

  10. Aparicio, T.: Resilent.js (2018). 16 March 2018


  8. Stagemonitor (2018). 17 April 2018

  complexity, modifiability, testability, maintainability and scalability were outlined for this purpose and represent the experience that might be taken into consideration. Finally, the practical application strengthens and supports decoupling, can help in development to focus on the essential moreover, increases efficiency in operation, not least through good opportunities for scaling.

  7. Spring cloud netflix (2018). 16 March 2018

  6. Spring cloud config (2018). 16 March 2018

  5. (2018). 16 March 2018

  4. Kubernetes (2018). 17 April 2018

  3. Elk stack (2018). 16 March 2018

  2. Docker compose (2018). 16 March 2018

  1. Apache Kafka (2017). 16 March 2018


  We would like to take this opportunity to thank the partners of

the research project. The research project WINNER is funded by the Federal Ministry

for Economic Affairs and Energy of Germany under project number 01ME16002D.


  79–84, March 2014.

  Microservice Architecture Within In-House Infrastructures



18. Krylovskiy, A., Jahn, M., Patti, E.: Designing a smart city internet of things

platform with microservice architecture. In: 2015 3rd International Conference on Future Internet of Things and Cloud, pp. 25–30, August 2015.



19. Linthicum, D.: Enterprise Application Integration. Addison-Wesley Information

Technology Series. Addison-Wesley, Essex (2000)


20. Mar´echaux, J.L.: Combining service-oriented architecture and event-driven

architecture using an enterprise service bus. IBM Developer Works, pp. 1269–1275 (2006)

  21. Newman, S.: Building Microservices. O’Reilly (2015) measurement infrastructure. In: 2013 Joint Conference of the 23rd International Workshop on Software Measurement and the 8th International Conference on Software Process and Product Measurement, pp. 63–70, October 2013.


23. Vianden, M., Lichter, H., Steffens, A.: Experience on a microservice-based

reference architecture for measurement systems. In: 2014 21st Asia-Pacific Software Engineering Conference, vol. 1, pp. 183–190, December 2014.



24. Wu, M., Lu, T.J., Ling, F.Y., Sun, J., Du, H.Y.: Research on the architecture of

internet of things. In: 2010 3rd International Conference on Advanced Computer Theory and Engineering(ICACTE). vol. 5, pp. V5-484–V5-487 (Aug 2010).


25. Yang, Z., Yue, Y., Yang, Y., Peng, Y., Wang, X., Liu, W.: Study and application on

the architecture and key technologies for IoT. In: 2011 International Conference on Multimedia Technology, pp. 747–751, July 2011.



26. Zanella, A., Bui, N., Castellani, A., Vangelista, L., Zorzi, M.: Internet of things

for smart cities. IEEE Internet Things J. 1(1), 22–32 (2014).



27. Zimmermann, O.: Microservices tenets. Comput. Sci. Res. Dev. 32(3), 301–310



Multi-agent Architecture of a MIBES

for Smart Energy Management




  1 B 1 J´er´emie Bosom Anna Scius-Bertrand , Ha¨ı Tran , and Marc Bui

CHArt Laboratory EA 4004, EPHE, PSL Research University, Paris, France


2 Energisme, Boulogne-Billancourt, France


  This paper introduces the concept of Multi-Institution

Building Energy System (MIBES) for the smart energy management.

The MIBES addresses the exploitation of energy data shared by numer-

ous multi-site multi-purpose institutions. It is a “hierarchical graph”

describing the physical and structural reality of the data collected for

these institutions. We propose the architecture of a multi-agent system

(MAS) for the MIBES smart management. This MAS is then used within

a data collection system to allow real time treatment of the system. This

complete system is being deployed in a french company called Energisme.

  Keywords: Multi-agent system · Data collection system Building energy modeling

1 Introduction

  During Kyoto Protocol’s signature in 1997, a large number of countries commit- ted themselves to reduce their gas emissions by 5% of the record in 1990. Since then, buildings have been identified as the biggest energy consumer sector for European countries.

  New laws and regulations are adopted in numerous countries to enforce sus- tainable energetic performances in post-production building life cycle. In France, some laws as “Loi Grenelle” I and II intimate public agencies and advice com- panies to take actions in order to reduce their energy consumptions.

  These new regulations accentuate the need of knowledge and control over energy consumptions for both private and public institutions. Nevertheless, the smart management of energy consumption data is a complex task due to the interdependence of energetic factors within buildings

   ]. The task is even more

  challenging when it involves Big Data ].

  For these reasons, institutions call the services of third party companies spe- cialized in on-site energy sensors installation and in energy data management.


Multi-agent Architecture of a MIBES for Smart Energy Management


  This work takes place in one these companies called Energisme

   ]. The set of

  sites managed by this kind of companies has the following properties:

  • – Sites repartition is not contiguous: data are gathered from various clients institutions with extremely different activities spread over several sites and countries,
  • – The possible actions on the sites are very limited or non-existent: these third- party companies do not take on direct interventions to reduce energy con- sumption as can be seen for building automation,
  • – The data heterogeneity and volume impose the use of Big Data techniques: data are gathered from smart meters, energetic invoices in PDF format pro- cessed automatically, industrial processes. . .

  Despite a through review of the literature, we have not found a defined sys- tem with the exact same requirements. Indeed, the main topics related to this kind of systems are more focused on electricity and spatial relations, such as Smart Grids, or on smart meters management and control. This paper suggests a definition and a solution for the smart management of such a system in the form of a hierarchical graph called Multi-Institution Building Energy System (MIBES). The MIBES reflects the physical and structural reality of the studied system.

  The contribution of this article is the architecture of a Data Collection System


] (DCS) for the constitution of a database describing a MIBES. This DCS shall

  ensure the data collection and analysis while meeting the requirements for Big Data management.

  The DCS subsystem in charge of the MIBES smart management is a Multi- Agent System (MAS). It allows concurrent and autonomous simulation of the various hierarchical elements defined by the MIBES.

  This paper is organized as follows: Sect.

   defines the MIBES and its dif-

  ferences with already existing concepts. Section

   presents the MIBES formal

  definition as a hierarchical graph and its architecture. Section

   defines a proper

  Big Data architecture for a DCS and the architecture of its MAS in charge of the MIBES modeling.

2 The MIBES Concept for Smart Energy Management

  We introduce the concept of Multi-Institution Building Energy System (MIBES) as the data shared by several institutions in order to manage their energy con- sumptions.

  Institutions decide on data storage and data exploitation policies. Therefore, every retrieved data has to be labeled, stored and then used accordingly with its owner policy. Consequently, all data can not be shared with the entire system. And when data are retrieved from a same one building used by several insti- tutions, the policies of all the institutions involved need to be considered. To illustrate this particularity of the studied system, the Fig.

   shows a common

20 J. Bosom et al.

  Fig. 1.

  One example of MIBES common case involving several institutions in a same building

  case involving several institutions that the MIBES should handle. This explains why our system is a “Multi-Institution” one.

  Most of entities embedding sensors are buildings. We restrict ourselves to “Building Energy System”. The MIBES is mainly based on data about build- ings characteristics, their installations or their use. Building modeling is a hard task due to the interdependence of energetic factors within buildings

   ]. The considered buildings are non-residential or collective residential.

  The reality of multiple policies for data exploitation, the discontinuous geo- graphical distribution of sites, the limitation of possible actions of the system, the heterogeneity and volume of the data and the inherent complexity of building modeling prompted us to search for existing smart management systems alike the MIBES in the literature.

  Multiple smart energy management systems exist such as the Cyber-Physical Systems

   ], the Sensor


  , the Multi-agent

  Based Energy Management Systems

   ], the Automated Production Systems


   ]. These systems mainly share the

  same similarities and differences as those listed in the following part between our solution, the Smart Grids

   ] and the Annex 53 of the International Energy

  Agency .

  The Smart Grid

  is a greatly studied smart energy management system. It is similar to the MIBES in that it manages a great number of multi-purpose sites.

  However, there are substantive differences between a MIBES and a Smart Grid residing both in their subject matter and in their objectives. The Smart Grid has as goal the optimization of the existing electrical grid. Therefore, it man- ages producer-consumers entities and aggregates their behavior by geographical


Multi-agent Architecture of a MIBES for Smart Energy Management


  proximity into entities such as Virtual Power Plants

   ]. This

  approach is irrelevant in our case because we manage entities spread over sparse localizations. Furthermore, we manage only energy consumers and we do not integrate questions of optimization between energy demand and supply.

  Nevertheless, we retain some aspects from the study of Smart Grids and we are specially interested on their resilience. The resilience of a system is its capacity to retrieve a stable state after suffering a disturbance. Our system has to be able to cope with:

  • – to new elements,
  • – to the heterogeneity of data,
  • – to faults, crashes and errors, – to potential attacks.

  Consequently, the required specification of our system is to be autonomic, that is to be able to manage its own operations. For this, our system needs to meet the requirements of the self-management as: (1) self-configuration and boot, (2) self-healing, (3) self-optimization, (4) self-monitoring, (5) self-protection.

  In the Annex 53

  , the International Energy Agency identifies six main

  factors influencing energy consumption in buildings. Also, the Annex 53 specifies that, in order to evaluate building energy performances, a suitable database must first be made available. This database needs to be aggregated from the available buildings data. The Annex 53 suggests for that the use of a Data Collection System (DCS). A DCS is a system that collects data and aggregates them for their analysis and exploitation.

  Until then, we have presented the MIBES as a model relating the physical and structural reality of the studied system. However, the exploitation of the MIBES data requires the use of a suitable system performing data collection and analysis. For this reason, we will use a DCS as defined above. This DCS has several challenges to meet:

  • – allowing Big Data treatments (scalable, . . . ),
  • – the MIBES smart management: exploiting the different data sources and adapting the modeling to be as close as possible to what we want.

  The MIBES smart management part of the DCS needs to be structured accordingly to the requirements from the Smart Grid review we want to keep. To do so, we use a bottom-up solution in the form of a Multi-Agent System (MAS)

  . By the aggregation of data through levels of modeling, the bottom- up approach allows the construction of knowledge about the studied system.

  The multi-agent approach is particularly designed for distributed structures as a network of sensors. The agents of a multi-agent system are autonomous pro- grams that are meant to be executed in parallel independent processes. They are also adaptive and evolve according to the interactions they have with other agents. Furthermore, a MAS allows to manage new elements in a transparent and autonomous way.

  Now that we have framed the MIBES approach, we introduce a formal defi- nition of MIBES in the next section.

  22 J. Bosom et al.

  3 Multi-Institution Building Energy System (MIBES)

  The Multi-Institution Building Energy System (MIBES) is an acyclic bidirec- tional graph G with hierarchical levels defined as: G

  = (V, E) Where V is the set of vertices composed by the different elements interacting in the MIBES and E is the set of edges describing the links and communications between the elements of V .

  The MIBES separates the elements of V into a set H of four hierarchical levels aggregating data according to four criteria:

  • – by same physical object,
  • – by same space delimitation of the physical location,
  • – by same groups of sites that are defined for convenient management or sites comparisons, – by same owner institutions.

  Fig. 2.

  An example of a MIBES hierarchical graph

  We define S as the set listing the different types of possible elements in V . For each level of H, a corresponding type of elements is defined in S. An additional type of elements is added to manage the incoming data feeding the different levels of H with data. Therefore, the set S is composed of the five following types: (1) data sources, (2) physical assets, (3) sites, (4) building stocks and (5) institutions.

  We call V (s), s ∈ S the set of vertices in V and with type s. We specify that:

  • – At each node v ∈ V (Data Sources), data are emitted but must be labeled with an institution tag,
  • – At each node v ∈ V (s), s = Data Sources, an aggregate describing the state of the element is made available for the parents or neighbors of v.


Multi-agent Architecture of a MIBES for Smart Energy Management


  For better understanding, the Fig.

   shows an example of a MIBES graph

  and the following part gives a more detailed description of the MIBES elements and their roles.

  Data Sources:

  Data sources are the elements in charge of the streaming of incoming data. They constitute the base elements of our system and feed all other elements with their processed data. They are a flexible way to integrate incoming data at every level of the MIBES hierarchy by bringing:

  • – best possible qualified data in our possession,

  Physical Assets: We represent the physical assets as materials within a site.

  They are associated with at least one data source giving the available information about the on-site installation of the physical asset. This category aims to regroup gateways, sensors and technical processes like boilers that may emit or not data depending on how old the process is.


Sites: The concept of buildings inside MIBES is what we call a site. Depending

  on institutions activities, sites could be a floor of a building, an entire building, several ones and it could even include an additional land.


Building Stocks: Stocks that are defined as a set of sites fulfilling the same

  function. This definition is willingly large to encompass all possible building stock sizes and make all wanted levels of management easier. Thus, the richer in data a set of sites is, the more it can be cut into a precise hierarchy of building stocks.


  Last level of the MIBES hierarchy, the institutions are public or private entities managing non-residential or collective residential building stocks. Most of available data can be assimilated to time series because new data are produced over time. Consequently, we can use state space modeling to construct aggregates from available time series

   ]. These aggregates can be used as state representations for upper hierarchical levels or neighbors.

  The distributed database of aggregates described by the MIBES can be used as a case database. Therefore, any new element with enough data to compute an aggregate can be compared to the case database. This comparison allows to find similar elements and choose the more appropriate behavior to manage the new element.

  MIBES, as a hierarchical graph using five types of modeling elements inter- acting with each other, permits us to build a case database to compare elements and manage new incomers. In this context, our goal is to design a solution allow- ing the construction of a study case database so we can exploit and manage information collected through a MIBES over time.

  In order to achieve this goal, we define in the next part the architecture of a Data Collection System.

24 J. Bosom et al.

  4 A Data Collection System for MIBES Smart Management

  In this section, we first focus on the definition of a Data Collection System (DCS) architecture for Big Data treatments and MIBES smart management. Then, we present a Multi-Agent System (MAS) architecture as a DCS subsystem ensuring the smart management of the MIBES.

  4.1 Data Collection System Architecture The DCS is fed by a database, called Data Service, giving access to all data retrieved for the available institutions. In more detail, the collected data can come from:

  • – thousands of sensors of two main kinds distributed over hundreds of sites: passive sensors or gateways that are just relaying data from other sensors
    • to our servers, active sensors measuring all the available energies (electricity, water, gas,
    • heat flows, temperatures, etc.),

  • – data automatically extracted from tens of thousands of energetic bills in PDF format produced by a large variety of energy providers,
  • – technical processes as automated production lines or boilers,
  • – users,
  • – additional data sources, for example, data extracted from Customer Relation- ship Management (CRM) systems, – independent institutions providing weather data, etc.

  The system regroups data about non-residential or collective residential buildings stocks as: plants, hotels, public swimming pools, schools, supermar- kets, etc. For instance, little territorial communities are typically only sharing monthly electricity and water bills in PDF format about public buildings. In con- trast, hypermarket chains equip their sites with Internet of Things (IoT) sensors measuring hourly water and electricity consumptions and fridges temperatures. Organizations such as those managing asphalt mixing plants can give access to their technical processes logging production every 5 s.

  The smart management of this database needs an adapted solution for Big Data. Consequently, the DCS implements a LAMBDA-like architecture

   ] for

  real time and batch computing. The architecture separates hot and cold data treatments. As shown in Fig.

  the DCS uses a streaming system suited for Big Data to manage new data coming from the Data Service.

  The streaming system provides views of incoming data to a MAS in charge of the MIBES smart management. Hot treatments are done using the provided views. For cold treatments and experiments, data are stored in a local database.

  By knowing the MAS is in charge of the MIBES smart management, it needs to perform intense computations on available data as data mining and machine learning treatments. MAS is composed of asynchronous agents. Therefore allow- ing agents to execute large computations in an autonomous way can lead to


Multi-agent Architecture of a MIBES for Smart Energy Management

  25 Fig. 3.

  Data Collection System architecture

  several problems including system overload. This challenge is referred in the literature as Many-Task Computing


  We suggest to use a Task Queue Manager (TQM) in the DCS architecture as a solution to this Many Task Computing problem. A TQM is a tool to manage a queue of computations to execute. MAS large computations are provided in the form of tasks to execute on specified data retrieved directly from either the streaming system or the local database. The TQM dispatches the tasks and optimally use the available memory and computation power. It schedules tasks taking their priority, estimated execution time and resource requirements into account. Computations are done using specialized data processing software and machine learning libraries executed on Central Processing Unit (CPU) and Graphics Processing Unit (GPU).

  The DCS architecture refines MAS context and requirements and ensures the scalability of the system. The next part presents a MAS architecture for the specified DCS.

  4.2 Multi-agent System Architecture for MIBES Simulation As explained in Sect.

  the MIBES is a hierarchical graph of five types of ele-

  ments. The MAS architecture exploits these five elements as base elements named MIBES agents. In this section, we define more precisely these MIBES agents, and then present the whole architecture.

26 J. Bosom et al.

  MIBES Agents. We define an AbstractMIBESAgent as a base element of the five types of MIBES agents implementing the five different entity types interacting in MIBES.


  Each MIBES agent of the MAS has some latent vari- ables to predict. These variables are:

  • – a boolean defining whether the agent is currently in activity,
  • – a boolean for the error state of the agent,
  • – a boolean for the security state of the agent, – a state transition function to predict the next state from the current one.

  The state of an agent can be defined, as stated in Sect. t as an aggre- gate constructed on available data. Considering an aggregate a of an Abstract- MIBESAgent at time t, we can define a state transition function as: t t f (a ) = a

  • 1

  This state transition function can be approximated using Machine Learning (ML) techniques on the data history of the agent. Therefore, we can predict the future behavior of the agent.

  Fig. 4.

  DataSourceAgent modeling and examples of sub-classes

DataSourceAgent. This agent is the abstract class of all data sources. It has

one monitoring variable as a time series describing the incoming stream of data.

  Some classes inherit from the data sources as shown in Fig.



  We need to give information to the PhysicalAssetAgent about its own physical location and installation. The easiest way to do so is by using a DataSourceAgent dedicated to the streaming of such data. For this


Multi-agent Architecture of a MIBES for Smart Energy Management

  27 Fig. 5.

  PhysicalAssetAgent modeling and examples of sub-classes reason and as shown in Fig. the PhysicalAssetAgent is associated to one Data-

  SourceAgent. The latter indicates only physical characteristics of the agent that change a bit over time, like its position in the site. Assets are useful to model some black box components in a site as gateways, old boilers, etc.

  We also define a SensorAgent as a PhysicalAssetAgent fed by one or several DataSourceAgent(s). For instance, an energetic delivery point for electricity will be modeled as a subclass of SensorAgent. It will aggregate information from a lot of data sources corresponding to as many electricity bills.


SiteAgent. SiteAgent aggregates data coming mainly from DataSourceAgents

  and PhysicalAssetAgents. As specified in the MIBES, a SiteAgent could also be in a relationship of child/parent or neighbors with other SiteAgents.


  BuildingStockAgents interact with DataSourceAgents, SiteAgents and other BuildingStockAgents. These agents can be useful to aggre- gate datasets of some types of sites easily for monitoring and experiments.

  The splitting of a set of sites into a BuildingStockAgent hierarchy can be given by the owning institution or it can also be automatically or manually adjusted. For instance, we can consider automatic stock building splitting using clustering techniques for new sites without a building stock hierarchy, and more manual splitting for the most detailed and/or oldest sites. The possibility of having some automatic or manual cut-outs is an essential feature of our system. It allows both the automatic processing of the data and the possible interactions and parameterizations of the users of the system.


InstitutionAgent. The InstitutionAgent uses data from BuildingStockAgents

and can define neighbors or parent/child relations with other InstitutionAgents.

  It allows, among others, to study the influence of an institution’s management policy.

28 J. Bosom et al.

  Fig. 6.

  UML architecture of the agents System Agents for MIBES Agents Hierarchy.

  The whole hierarchy of the MIBES agents interacts with other agents in charge of the management of specific tasks. Thus, as shown in Fig.

  the AbstractMIBESAgent, and so the

  different levels of modeling, interacts with the following agents:

  • – DataAgent: agent in charge of data and internal memory,
  • – ErrorManagerAgent: agent in charge of the detection and management of

  ErrorEvent ,

  • – SecurityManagerAgent: agent responsible for detection and management of

  SecurityEvent ,

  • – InterventionManagerAgent: agent in charge of the actions needed to solve the detected ErrorEvent and SecurityEvent. It also keeps track of possible interventions on the site coming both from our own actions and from the site’s institutions,
  • – ApplicationManagerAgent: manager agent for specific sites’ applications as predicting a specific business process behavior, etc.
  • – AggregatorManagerAgent: agent responsible for constructing the aggregations of required data for the achievement of the AbstractMIBESAgent’s objectives AggregatorAgent.

  The AggregatorManagerAgent manages a list of Aggregator- Agents . As shown in Fig.

  each modeling level instantiates a specific subclass of

  AggregatorAgent to aggregate their data. The Fig.

   shows an implementation of

  building energy modeling influence factors as described in

  . In this case, one

  agent is in charge to aggregate information about one of the factors using one or multiple models trained on a large instance of sites.

  Computation Model. As stated in Sect. agents expensive computations

  are managed by a Task Queue Manager (TQM). Therefore, our MAS needs a


Multi-agent Architecture of a MIBES for Smart Energy Management

  29 Fig. 7.

  UML architecture for AggregatorAgent

  computation manager as an interface with the TQM. For this task, we will define a GlobalComputationSupervisorAgent, as shown in Fig.



Algorithms Managing. As previously stated, the MAS system needs to use

  ML algorithms for its predictive parts. The most suited algorithm depends on the prediction goal and on the available data. However, the volume and heterogeneity of MIBES data raise challenges.


Fig. 8. UML for the whole multi-agent architecture

  To overcome these challenges, we maintain a library of algorithms. Each algorithm defines requirements on input data for training and testing. This way, algorithms can be executed easily on any given data meeting the requirements. Furthermore, this library can be enriched over time by the users of the system.

30 J. Bosom et al.

  As described in Sect.

  the MIBES allows us to determine the best behavior

  to handle new incoming cases. The library of algorithms helps achieving this goal by providing a unified view of the possible treatments on data. This way, it is possible to evaluate the available algorithms on each known cases of the MIBES case database and so to determine what treatments are the best for new cases.

  4.3 Multi-agent System Global Architecture The MAS is separated into several multi-agent subsystems specialized in different tasks, as shown in Fig.


  The core part of the system is dedicated to the execution of the following agents:

  • – GlobalErrorSupervisorAgent: manages errors at the system level
  • – GlobalSecuritySupervisorAgent: manages security at the system level
  • – GlobalInterventionSupervisorAgent: manages interventions at the system level and regulates the flow of mails sent to IoT technicians and managers of the system to schedule interventions,
  • – GlobalAlgorithmSupervisorAgent: manages the algorithms library
  • – GlobalComputationSupervisorAgent: manages the computation flow of agents
  • – GlobalExperimentSupervisorAgent: manages experiments created by searchers and managers of the system.
  • – GlobalMonitoringSupervisorAgent: manages the self-monitoring of the MAS

  To this subsystem, the MAS MIBES Simulation process launching and managing the multi-agent simulation of the whole MIBES agent architecture described in Sect.

   is added.

  A subsystem is also dedicated to the self-monitoring of the MIBES simula- tion. The self-monitoring is needed to achieve sustainable self-management as required in Sect.


  The MAS API makes the link between the MAS and the external world, represented by the Data Service. By separating the global managers from the rest of the system, we can control the communications received by these processes more precisely. The same is true for the two other subsystems. It allows to separate critical points, computations and requests-responses flow from the outside world. This helps make the system easier to secure and maintain.


  In this paper we have addressed the problem of the smart management of build- ing energy systems for multi-purpose multi-site institutions. We have introduced the concept of Multi-Institution Building Energy System (MIBES) to define the physical and structural reality of this system. The MIBES, as a hierarchi- cal graph, defines several hierarchical levels for the smart management as data sources, physical assets, sites, building stocks and institutions. By interacting


Multi-agent Architecture of a MIBES for Smart Energy Management


  with each other using aggregates, layers allow to create a complex view of the system as precise as one might wish.

  Then, we have presented the architecture of a Data Collection System (DCS) suited for Big Data treatments and MIBES smart management. This last part is ensured by a Multi-Agent System (MAS) for which we suggest an architecture. The MAS uses subsystems managing security, errors, interventions, computa- tions, experiments and algorithms library. This allows real time supervision and efficient smart management.

  The undergoing work is addressing the DCS elaboration and the benchmark- ing of the MIBES database.


Acknowledgement. This work has been funded by the french company Energisme

. Related and future works are currently being held as projects within Energisme.


1. EBC Annex 53 Total Energy Use in Buildings: Analysis & Evaluation Methods.

  2. Energisme. Accessed 24 April 2018


3. Ahat, M., Amor, S.B., Bui, M., Bui, A., Gu´erard, G., Petermann, C.: Smart grid

and optimization. Am. J. Oper. Res. 3(1A), 196–206 (2013)


4. Aoki, M.: State Space Modeling of Time Series. Universitext. Springer,

Heidelberg (2013).


5. Gazafroudi, A.S., De Paz, J.F., Prieto-Castrillo, F., Villarrubia, G., Talari, S.,

Shafie-khah, M., Catal˜ ao, J.P.S.: A review of multi-agent based energy management systems. In: De Paz, J.F., Juli´ an, V., Villarrubia, G., Marreiros, G., Novais, P. (eds.) ISAmI 2017. AISC, vol. 615, pp. 203–209. Springer, Cham (2017).


6. Groen, F.C.A., Spaan, M.T.J., Kok, J.R., Pavlin, G.: Real world multi-agent sys-

tems: information sharing, coordination and planning. In: ten Cate, B.D., Zee- vat, H.W. (eds.) TbiLLC 2005. LNCS (LNAI), vol. 4363, pp. 154–165. Springer, Heidelberg (2007).


7. Haynes, B., Nunnington, N., Eccles, T.: Corporate Real Estate Asset Management:

Strategy and Implementation. Taylor & Francis, London (2017). google-Books-ID: xCQlDwAAQBAJ


8. Jirkovsk` y, V., Obitko, M., Maˇr´ık, V.: Understanding data heterogeneity in the

context of cyber-physical systems integration. IEEE Trans. Industr. Inf. 13(2), 660–667 (2017)


9. Karthick, A., Ramaraj, E., Subramanian, R.G.: An efficient multi queue job

scheduling for cloud computing. In: 2014 World Congress on Computing and Com- munication Technologies (WCCCT), pp. 164–166. IEEE (2014)


10. Kavgic, M., Mavrogianni, A., Mumovic, D., Summerfield, A., Stevanovic, Z.,

Djurovic-Petrovic, M.: A review of bottom-up building stock models for energy consumption in the residential sector. Build. Environ. 45(7), 1683–1697 (2010).

32 J. Bosom et al.


11. Lukovic, S., Kaitovic, I., Mura, M., Bondi, U.: Virtual power plant as a bridge

between distributed energy resources and smart grid. In: 2010 43rd Hawaii Inter- national Conference on System Sciences (HICSS), pp. 1–8. IEEE (2010).


12. Lunacek, M., Braden, J., Hauser, T.: The scaling of many-task computing

approaches in python on cluster supercomputers. In: 2013 IEEE International Con- ference on Cluster Computing (CLUSTER), pp. 1–8. IEEE (2013)


13. Surmann, M., Brunauer, W.A., Bienert, S.: The energy efficiency of corporate

real estate assets: the role of energy management for corporate environmen- tal performance. J. Corp. Real Estate 18(2), 68–101 (2016).


14. Marz, N., Warren, J.: Big Data: Principles and Best Practices of Scalable Realtime

Data Systems. Manning Publications Co., Shelter Island (2015)


15. Prabhu, B., Balakumar, N., Antony, A.J.: Wireless Sensor Network Based

Smart Environment Applications. SSRN Scholarly Paper ID 2909105, Social Sci- ence Research Network, Rochester, NY, January 2017.


16. Saad, W., Han, Z., Poor, H., Basar, T.: Game-theoretic methods for the smart

grid: an overview of microgrid systems, demand-side management, and smart grid communications. IEEE Sig. Process. Mag. 29(5), 86–105 (2012).


17. Trunzer, E., Kirchen, I., Folmer, J., Koltun, G., Vogel-Heuser, B.: A flexible archi-

tecture for data mining from heterogeneous data sources in automated production systems. In: 2017 IEEE International Conference on Industrial Technology (ICIT), pp. 1106–1111. IEEE (2017)


18. Zhao, F.: Agent-based modeling of commercial building stocks for energy

policy and demand response analysis. Georgia Institute of Technology (2012).


19. Zhou, K., Fu, C., Yang, S.: Big data driven smart energy management: from big

data to big insights. Renew. Sustain. Energy Rev. 56, 215–225 (2016).


A C-ITS Central Station

as a Communication Manager

  Geoffrey Wilhelm, Hac`ene Fouchal

  ( B



  , Kevin Thomas, and Marwane Ayaida


CReSTIC, Universit´e de Reims Champagne-Ardenne, Reims, France


Marwane.Ayaida }

Abstract. A C-ITS is a system where mobile stations OBU (On-Board


Units) exchange messages with other ITSS-V or RSU (Road Side Units).

Messages are sent through a specific WIFI (IEEE 802.11p) denoted also

ETSI ITS-G5. The efficiency of this technology has been proven in terms

of latency. However, RSU are common everywhere, for this reason we look

for another mean to guarantee this communication. Cellular networks are

widely deployed and ma support these communications.


In this paper, we present an architecture which ensures end-to-end

communication between mobile nodes which integrates security aspects

in particular authenticity of exchanged messages. The privacy is also



We have measured some indicators as latency (notification delay),

packet delivery ratio (number of messages arrived after a threshold).

These indicators confirmed that our proposed architecture has a inter-

esting performances and could be deployed widely.

  Keywords: C-ITS ·

  VANETs ·

  Cellular networks Hybrid communications

1 Introduction The deployment of connected vehicles is an interesting challenge since a decade.

  The connectivity is one of the most important issue to solve. Indeed, a dedi- cated WIFI has been designed for connected vehicles: IEEE 802.11p (denoted also ETSI ITS-G5). However, the deployment of ITS-G5 hotspots (denoted Road Side Units) s not generalised. This deployment of such technology takes a lot of time and is an expensive task. Indeed, the penetration rate of the connected vehicles is increasing slowly. Therefore, the coverage of such technology remains limited. However, it is very important to receive the events to avoid accidents and save lives. To deal with this, the coverage could be enhanced using the cellu- lar communication. In this paper, we intend to use the cellular network (3G/4G) in order to ensure the collection and the delivery of warning messages to and from vehicles. Every vehicles send continuously it Cooperative Awareness Mes- sages (CAM) to the Central ITS Station (ITSS-C). The latter maintains the


G. Wilhelm et al.

  location of the vehicles up-to-date. If an event is triggered in a zone, the event is then automatically forwarded using cellular communication to the nodes that are in the relevance area. The remainder of this paper is organised as follows: Sect.

   details the architecture of the pro-

  posed system. Section

   concludes the paper and gives some hints about future works.

2 Related Works



] proposes an evaluation of vehicular communications networks through car

  sharing scenarios. The authors have investigated three parameters. They adopted a specific mobility model which has been imported to a simulator. They have worked on a grid Manhattan network and they observed some performance parameters such as delay, packet loss, etc. The most important objective of the study is to show that vehicular communication is feasible and realistic under some conditions.


   ] studies throughput over VANETs system along an unidirectional traffic

  for different conditions and transmission ranges of wireless equipments. All stud- ied vehicles are randomly connected. The paper gives few results of simulation studies achieved on NS-2 toolbox. They have measured performances indicators in case of congestion. A comparison of the obtained results with the expected connectivity has been done and have shown that the throughput over simulation is lower due to packet losses caused by collisions.

  Authors of

  presents an alternative to WAVE/DSRC using an hybrid

  system, which uses Wi-Fi Direct and Cellular Network. They show that such a system could work for C-ITS. However, this paper does not take into account the hybridation between ITS-G5 and Cellular Network.


   ] presents another alternative to WAVE/DSRC solution using here Wi-Fi

  Direct, ZigBee and Cellular Network. Wi-Fi Direct is used as a direct link between nodes. ZigBee is used to connect roadside sensors and Cellular Net- work for long distance communication. In this study, the ITS-G5 is also ignored.


   ], the authors provide their network architecture which has been deployed

  in Spain, where communicating vehicles are switching between 802.11p and 3G, depending on RSU’s availability.


   ] presents a detailed study on performance evaluation of IEEE 80211.p net-

  works versus LTE vehicular networks. The authors analyzed some performance indicators like the end-to-end delay for both networks in different scenarios (high density, urban environments, etc.). Many important issues have been measured as network availability and reliability. The authors have proved through simula- tions that LTE solution meets most of the application requirements in terms of reliability, scalability, and mobility. However, IEEE 802.11p provides acceptable performance for sparse network topologies with limited mobility support.


   ] gives an efficient solution for routing messages over VANETs by using the vehicle’s heading.


   ] gives an overview of how research on vehicular communication evolved in

  Europe and, especially, in Germany. They describe the German field operational

  A C-ITS Central Station as a Communication Manager


  test sim TD. The project sim TD is the first field operational test that evaluated the effectiveness and benefits of applications based on vehicular communication in a setup that is representative for a realistic deployment environment. It is, therefore, the next necessary step to prepare for an informed deployment decision of cooperative systems.



  is a study about movement prediction of vehicles. Indeed, an adapted routing algorithms are proposed in

   ] gives an overview of strategies

  to use for routing on VANETs.

   ] reviews much more actual strategies on vehicular networks.

  All the works presented below handle the communications between vehicles using cellular networks. In this study we show how to handle both ETSI ITS-G5 and cellular networks for communication between vehicles.

3 System Architecture

  3.1 Preliminaries In the area of C-ITS (Cooperative Intelligent Transportation System), a protocol stack has been defined and standardised by the ETSI standardisation institute in Europe

   ]. Over the Transport-Networking layer (defined as geo-networking


   ], the Facilities layer has been designed in order to be an efficient inter-

  face between the application layer (close to the driver and the vehicle sensors) and the Transport-Networking layer

   ]. The Facilities layer handles 2 main

  messages: CAM (Cooperative Awareness Message)

  and DENM (Decentral-

  ized Environmental Notification Message)

  . The aim behind sending CAM

  messages is to give dynamic information about the vehicle (i.e. position, speed, heading, etc.). The aim of exchanging DENM messages is to notify about any type of event (i.e. accident, traffic jam, etc.). Whenever the vehicle moves, it sends CAMs to other vehicles using Vehicle-to-Vehicle (V2V) communications. Depending on the speed of the vehicle, the frequency of CAM messages changes from 1 Hz to 10 Hz.

  3.2 General System Architecture Description Usually, CAMs and DENMs are exchanged through ETSI ITS-G5 communi- cation channel. In our solution, these two types of messages will be sent also through the cellular medium. However, ETSI ITS-G5 is a broadcast-based chan- nel and cellular networks is a unicast-based channel and then will use end-to- end connections through TCP connections. For this end, we need to specify the behaviour of both the mobile node (ITSS-V) as well as the central station (ITSS-C). Figure

   shows the global architecture of the whole system. In fact,

  whenever a vehicle is moving, it sends its messages either CAMs or DENMs through ETSI ITS-G5 and duplicated through cellular networks. When the cen- tral station (ITSS-C) receives for example a CAM from a vehicle, its position is


G. Wilhelm et al.

  updated on the vehicle location data base. When it receives a DENM, it stores it in the DENM database. In this scheme depicted by the Fig.

  we have two


  • – vehicle side: the vehicle works in the same network environment than for the ITS-G5 one.
  • – central station side: this station handles the CAM messages for updating the vehicle’s position and the DENM messages through some tasks that will be detailed below.


Fig. 1. The general Central Station Architecture

  3.3 Central Station Description This subsection is dedicated to the description of the central station (ITSS-C) component. This node is responsible of almost two tasks: (1) Location manage- ment and (2) Event management. These tasks will be presented below.

  A C-ITS Central Station as a Communication Manager



   details the components and the tasks handled by the central station


  1. Global Database: this database, located at the central station (ITSS-C), is responsible of storing all DENMs sent by all the connected vehicles.

  2. A process extracts the interesting DENMs for a specific vehicle and copy them to a temporary database denoted Local Database used by the Mobile Node Manager.

  3. Local Database: this database, located at the Mobile Node Manager, is respon- sible of storing the DENMs that concern the vehicle.

  4. All DENMs must be checked by the routine, if a Validity Duration is expired, it will be erased.

  5. DENMs of the local database are collected and sent to the vehicle using the cellular network.

  6. Vehicle receives DENMs through to a cellular connection and displays them top the driver at the right time and location.


Fig. 2. The components and tasks within the Central Station

  Location Management. The Algorithm

   describes the location management

  policy within the central station (ITSS-C). When it receives a DENM from an

  ITS station (ITSS-V), it extracts the different points of the Destination Area


  (i.e. North, East, South and West) using the OpenStreetMap AP . Then, we look for the zoom that covers the 4 sides. When, it is found the DENM is stored in the database of the central station (ITSS-C). 1


G. Wilhelm et al.

  Algorithm 1. Location management algorithm in the central station (ITSS-C)

  1: while (RECEIV E(denm)) do N orthP oint 2: ← getN orthOf DestinationArea(denm); EastP oint 3: ← getEastOf DestinationArea(denm); SouthP oint 4: ← getSouthOf DestinationArea(denm); W estP oint 5: ← getW estOf DestinationArea(denm); 6: for (zoom = 18; zoom < 0; zoom − −) do

  N orthOSM 7: ← OsmP osF romGP SP oint( N orthP oint, zoom 8: );

  9: EastOSM ← OsmP osF romGP SP oint( 10: EastP oint, zoom ); 11: SouthOSM ← OsmP osF romGP SP oint( 12: SouthP oint, zoom ); 13: W estOSM ← OsmP osF romGP SP oint( 14: W estP oint, zoom ); 15: if (N orthOSM == EastOSM == SouthOSM == W estOSM ) then 16: Store (zoom, N orthOSM, denm);

  U pdateServerV ersionN umber 17: (); 18: end if 19: end for 20: end while Event Management.

  The Algorithm

   describes the event management policy

  within the central station (ITSS-C). This algorithm mainly describes the Store procedure. When there is a DENM to be stored, we verify if the DENM is a new one. If it is the case, we add it in the Database. Otherwise, we update it in the storage if the referenceTime of the one received is higher than the stored one (i.e. it is a new update for an existing event).

  Algorithm 2. Event management algorithm in the central station (ITSS-C)

  1: procedure Store(zoom, OSM P os, denm) storedDenm 2: ← getF romStorage( 3: zoom, OSM P os, denm.actionId ); 4: if (storedDenm == null) then 5: addT oStorage (zoom, OSM P os, denm); 6: else if (storedDenm.ref erenceT ime > denm.ref erenceT ime) then 7: updateInStorage (zoom, OSM P os, denm); 8: end if 9: end procedure

  3.4 Client Description This subsection is dedicated to the description of the vehicle (ITSS-V) compo- nent. The vehicle is responsible of almost two tasks: (1) Location update and (2) Event notification. These tasks will be presented below.

  A C-ITS Central Station as a Communication Manager

  11: end if 12: if (routine) then 13: triggerEventN otif icationRoutine (); 14: end if 15: end while Event Notification.

  We have deployed an application over smartphones into vehicles. It collects data from vehicles. We have experimented the data collection during one week with

  1: procedure triggerEventNotificationRoutine( ) 2: for (osmP os ∈ currentP os) do 3: for (denm ∈ localStorage.get(osmP os)) do 4: if (!ackDatabase.contains(denm)) then 5: SEN D (denm); 6: end if 7: end for 8: end for 9: end procedure

  If it is the case, it sends it directly using the cellular communication. Algorithm 4. Event notification algorithm in the ITS station (ITSS-V)

  This procedure scans all the DENMs and verifies if it was not acknowledged to the central station (ITSS-C).

  within the ITS station (ITSS-V). It details the triggerEventNotificationRou- tine procedure called in the previous Algorithm

   describes the event notification policy

  The Algorithm

  7: end if 8: if (newP os! = currentP os) then 9: currentP os ← newP os; 10: routine = true;


  5: updateM yLocalStorage


6: routine = true;

  3: routine = f alse; 4: if (myV ersionN umber < serverV ersionN umber) then

  1: while (SEN D(cam)) do 2: newP os ← getOSM P os(cam);

  Algorithm 3. Location update algorithm in the ITS station (ITSS-V)

  within the ITS station (ITSS-V). When a vehicle is sending a CAM, it veri- fies if its VersionNumber is not the same one than the server or if its position has changed, than he launches a routine to ask for a new notification event if it exists. Otherwise, it waits for one of these two conditions to occur.

   describes the location update policy

  Location Update. The Algorithm

4 Evaluation and Performance Analysis


G. Wilhelm et al.

  20 vehicles. We have observed the behavior of these vehicles. Then, we have mod- eled the vehicle behavior and we have introduced this behavior into a simulator. After that, we have simulated a network with 200 vehicles and we have observed some indicators such as packet delay. In order to show the impact of the delay, we measure the average packets delay using HTTP and TCP connections. Figure

   shows the TCP one. The delay is measured as

  the duration between the event is triggering in a mobile OBU (On-Board Unit) and the reception of this event by all the other connected OBUs.

  First, we studied the delay of the TCP connection using some packets as depicted in the Fig.

   with a real experimentation.


Fig. 3. TCP delay

The delay presented by Fig. is between 244 ms and 258 ms. These promising

  delays drive us to deep this study by proposing larger simulations using 200 vehicles that will be described below.


Fig. 4. HTTP average delay

The average HTTP delay presented in Fig. is around 400 ms. However, Fig.


  depicts an average TCP delay of 250 ms. This difference could be explained by the processing overhead of the HTTP protocol, which is an application layer

  A C-ITS Central Station as a Communication Manager

  41 Fig. 5. TCP average delay

  protocol (i.e. layer 7 in OSI model) compared with the TCP protocol, which is a transport layer protocol (i.e. layer 4 in OSI model).

  As a consequence of this evaluation, the delay of the cellular communication for such road application (i.e. road works and hazardous warning) is satisfac- tory. Even with HTTP connections, the delay seems reasonable. Besides, the deployed central station (ITSS-C) will use TCP connection to exchange DENM notification.

5 Conclusion

  In this paper we show a hybridation system architecture. This system is based principally on a central station (ITSS-C) and several connected mobile ITS sta- tions (ITSS-V). Each ITSS-V sends its CAMs to the ITSS-C. This allows the

  ITSS-C to have an accurate precision of the location of each ITSS-V. When a DENM is triggered, it will be also transmitted to the ITSS-C. The latter com- putes a list of interested ITSS-V and forwards to them the event. In this way, the mobile stations are informed with an average delay of 250 ms of the event. This allows to counterbalance the lack of ETSI-ITS G5 lack of coverage, while ensuring good performances.

  As a future works, we intend to test better the scalability of our system with a large deployment. Then, we intend to use it as a base to add some more complex and interesting tasks like the Vehicular Cloud Computing (VCC).


Acknowledgement. This work was made possible by EC Grant No. INEA/CEF/

TRAN/A2014/1042281 from the INEA Agency for the SCOOP project. The statements

made herein are solely the responsibility of the authors.


G. Wilhelm et al.



1. European Telecommunications Standards Institute (ETSI).


2. IEEE Draft Standard for Amendment to Standard [for] Information Technology-

Telecommunications and information exchange between systems-Local and Metropolitan networks-Specific requirements-Part II: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) specifications-Amendment 6: Wireless Access in VehicularEnvironments, in IEEE Std P802.11p/D11.0 April 2010, pp. 1–35, 15 June 2010


3. Intelligent Transport Systems (ITS); Vehicular Communications; GeoNetworking;

  Part 4: Geographical addressing and forwarding for point-to-point and point-to- multipoint communications; Sub-part 1: Media-Independent Functionality. ETSI EN 302 636-4-1 V1.2.1, July 2014


4. Intelligent Transport Systems (ITS); Vehicular Communications; GeoNetworking;

  Part 5: Transport Protocols; Sub-part 1: Basic Transport Protocol. ETSI EN 302 636-5-1 V1.2.1, August 2014


5. Intelligent Transport Systems (ITS); Vehicular Communications; Basic Set of

Applications; Part 2: Specification of Cooperative Awareness Basic Service. ETSI EN 302 637–2 v. 1.3.2, November 2014


6. Weia, C.: V2X communication in Europe: From research projects towards stan-

dardisation and field testing of vehicle communication technology. Comput. Netw.

  55(14), 3103–3119 (2011)


7. Santa, J., Fernandez, P.J., Garcia, F.: Deployment of vehicular networks in high-

ways using technologies 802.11p and IPv6 technologies. I JAHUC 24(1/2)(1), 33–48 (2017)


8. Intelligent Transport Systems (ITS); Vehicular Communications; Basic Set of

Applications; Part 3: Specifications of Decentralized Environmental Notification Basic Service. ETSI EN 302 637–3 V1.2.2, November 2014


9. Lochert, C., Hartenstein, H., Tian, J., Fussler, H., Hermann, D., Mauve, M.: A

routing strategy for vehicular ad hoc networks in city environments. In: IEEE

  IV2003 Intelligent Vehicles Symposium. Proceedings (Cat. No.03TH8683), pp. 156– 161, June 2003


10. Ayaida, M., Barhoumi, M., Fouchal, H., Ghamri-Doudane, Y., Afilal Phrhls, L.:

A movement-prediction-based joint routing and hierarchical location service for

  VANETs. In: IEEE International Conference on Communications (ICC), Budapest, Hungary, pp. 1424–1428, May 2013


11. Ayaida, M., Barhoumi, M., Fouchal, H., Ghamri-Doudane, Y., Afilal, L.: HHLS:

a hybrid routing technique for VANETs Global Communications Conference (GLOBECOM), Anaheim, pp. 44–48. IEEE, December 2012


12. Lu, W., Bao, Y., Sun, X., Wang, Z.: Performance evaluation of inter-vehicle com-

munication in a unidirectional dynamic traffic flow with shockwave. In: Proceedings of the International Conference on Ultra Modern Telecommunications, ICUMT 2009, 12–14 October 2009, St. Petersburg, Russia, pp. 1–6 (2009)


13. Lu, W., Han, L.D., Cherry, C.R.: Evaluation of vehicular communication networks

in a car sharing system. Int. J. Intell. Transp. Syst. Res. 11(3), 113–119 (2013)


14. Menouar, H., Lenardi, M., Filali, F.: A movement prediction-based routing pro-

tocol for vehicle-to-vehicle communications. In: V2VCOM 2005, 1st International Vehicle-to-Vehicle Communications Workshop, co-located with MobiQuitous 2005,

21 July 2005, San Diego, USA, July 2005

  A C-ITS Central Station as a Communication Manager



15. Mir, Z.H., Filali, F.: LTE and IEEE 802.11p for vehicular networking: a perfor-

mance evaluation. EURASIP J. Wirel. Commun. Networking 2014, 89 (2014)


16. Seet, B.-C., Liu, G., Lee, B.-S., Foh, C.-H., Wong, K.-J., Lee, K.-K.: A-STAR:

a mobile ad hoc routing strategy for metropolis vehicular communications. In: Mitrou, N., Kontovasilis, K., Rouskas, G.N., Iliadis, I., Merakos, L. (eds.) NET- WORKING 2004. LNCS, vol. 3042, pp. 989–999. Springer, Heidelberg (2004).


17. Taleb, T., Ochi, M., Jamalipour, A., Kato, N., Nemoto, Y.: An efficient vehicle-

heading based routing protocol for VANET networks. In: Wireless Communications and Networking Conference 2006, WCNC 2006, vol. 4, pp. 2199–2204. IEEE, April


18. Zeadally, S., Hunt, R., Chen, Y.-S., Irwin, A., Hassan, A.: Vehicular ad hoc net-

works (VANETs): status, results, and challenges. Telecommun. Syst. 50(4), 217– 241 (2012)


19. Jeong, S., Baek, Y., Son, S.H.: A hybrid V2X system for safety-critical applica-

tions in VANET. In: 2016 IEEE 4th International Conference on Cyber-Physical Systems, Networks, and Applications (CPSNA), Nagoya, pp. 13–18 (2016)

20. Bhover, S.U., Tugashetti, A., Rashinkar, P.: V2X communication protocol in

  VANET for co-operative intelligent transportation system. In: International Con- ference on Innovative Mechanisms for Industry Applications (ICIMIA), Bangalore, pp. 602–607 (2017)

  Data Analytics and Models


Dynamic Social Network Analysis

Using Author-Topic Model

1,2( ) 1,3

  1 B 1 Kim Thoa Ho , Quang Vu Bui , and Marc Bui

CHArt Laboratory EA 4004, EPHE, PSL Research University, Paris, France


3 University of Education, Hue University, Hue, Vietnam

University of Sciences, Hue University, Hue, Vietnam


Abstract. In this paper, we proposed an agent-based model for ana-

lyzing dynamic social network associated to textual information using

author-topic model, namely Textual-ABM. Author-topic model is chosen

to estimate topic’s distribution transformation of agents in the agent-

based model since it models the content of documents and also inter-

ests of authors. Textual-ABM can be utilized to discover dynamic of a

social network which includes not only network structure but also node’s

properties over time. Furthermore, we introduced independent cascade

model based on homophily, namely H-IC. The infected probability asso-

ciated with each edge is homophily or similarity which measured based

on topic’s distribution. We have applied our methodology to a collected

data set from NIPS and have obtained satisfactory results.

  Keywords: Social network Dynamic network Diffusion · ·


Independent cascade model Agent-based model Author-topic model

· ·

1 Introduction

  Social networking research has attracted a lot of attention of researchers with appearance of science fields including social network analysis (SNA)

   ], com-

  munity detection

  and so on. However, in reality, social networks are always

  in a state of fluctuations which are difficult to model. Therefore, the analytical tendencies shift from the study of the static to the analysis of the dynamic of the social network. Recently, there are two major approaches for analyzing the dynamic concept in social networks which comprise the fluctuation of structure and the characteristic variation of nodes over time. The first approach is the anal- ysis that a dynamic network has been considered as a cumulation of snapshot networks

   ]. In another hand, dynamic

  social network analyzes

  concentrated on exploiting aspect that properties of

  nodes may transform over time since they have the ability to learn and adapt to the interactive process instead of static nodes in SNA. Agent-based modeling is often used to explore how network evolve. In this study, we will analyze social networking dynamically with a combination of these two approaches.

  48 K. T. Ho et al.

  Majority research from two approaches above hardly mentioned about the content of messages among users since they often focus on structure

   ] or

  just propose an agent-based model for simulating dynamic social network with- out based on content

   ]. Therefore, in this study, we utilize textual content in

  the interactive process of users with the purpose of analyzing dynamic on both network’s structure and user’s interest. Recently, there are various methods for textual mining, for instance, Latent Dirichlet Allocation (LDA)

  , Author-

  Topic Model (ATM)

  , etc. We choose the author-topic model to define user’s

  interest since it simultaneously models the content of documents and the inter- ests of authors while LDA only considers a document as a mixture of probabilistic topics, not take into account author’s interests.

  In this paper, we construct an agent-based model for analyzing dynamic social network associated with the textual information, namely Textual-ABM. The dynamic of a social network is demonstrated through the fluctuation of Textual-ABM including agent’s network structure and agent’s characteristics. Additionally, we propose an independent cascade diffusion model that infected probability is based on homophily, namely H-IC. Independent cascade (IC) model is spreading model in which there is a probability of infection associ- ated with each edge. The probability P is the probability of u infecting v.


  This probability can be assigned based on frequency of interactions, geographic proximity, or historical infection traces. In this study, the infected probability between two users is measured by similarity or homophily based on textual information. Each infected node has the ability to infect its neighbor in the next time step based on the probability associated with that edge. Some experiments are implemented for H-IC on both static and dynamic co-author network. The results demonstrated that the effectiveness of H-IC outperforms comparison with random diffusion on the static network. Moreover, our results also illustrated the fluctuation of active percentage for the diffusion process on a dynamic network instead of attaining and remaining stable state on a static network.

  The structure of our paper is organized as follows: Sect.



  2 Preliminaries

  2.1 Agent-Based Model An agent-based model (ABM) is a class of computational models for simulating the actions and interactions of autonomous agents. ABM has been used in many fields including biology, ecology and social science

   ]. There are three major

  elements in ABM including agents, their environment, and mechanisms of inter- action among agents. Firstly, agents are heterogeneous entities which comprise diverse characteristics and behaviors. Secondly, agent’s environment is a space that plays responsibility for reflecting the structure of the overall system and

  Dynamic Social Network Analysis Using Author-Topic Model


  supplying agents their perceptions and enabling their actions. Thirdly, interac- tion is a form of information exchange among agents which resulted in perception and behavior. Particularly, the essence of an ABM is the dynamics of the global system emerges from the local interactions among its composing parts.

  2.2 Topic Modeling Latent Dirichlet Allocation (LDA). Latent Dirichlet Allocation(LDA)


  is a generative statistical model of a corpus. In LDA, each document is may be considered as a mixture of different topics and each topic is characterized by a probability distribution over a finite vocabulary of words. The generative model of LDA is described with the probabilistic graphical model in Fig.


  However, LDA only considers a document as a mixture of probabilistic topics, not take into account author’s interests. Therefore, author-topic model (ATM) is proposed since it simultaneously models the content of documents and the interests of authors.

  (a) LDA (b) ATM

Fig. 1. Text mining methods: LDA and ATM

  Author-Topic Model (ATM). Author-Topic model (ATM)

   ] is a genera-

  tive model represents each document with a mixture of topics, as in state-of-the- art approaches like LDA, and extends these approaches to author modeling by allowing the mixture weights for different topics to be determined by the authors of the document. The generative model of ATM is described with a graphical model in Fig.


  , proceeds as follows:

  1. For each author a = 1, ..., A choose θ a ∼ Dirichlet(α)

  For each topic t = 1,..T choose φ t ∼ Dirichlet(β)

  2. For each document d = 1,.., D

  2.1. Given the vector of authors a d

  2.2. For each word i = 1,...,N d

  2.2.1. Choose an author x di d ) ∼ Uniform(a

  2.2.2. Choose a topic z di x di ) ∼ Discrete(θ

  2.2.3. Choose a word w di z di ) ∼ Discrete(φ

50 K. T. Ho et al.

  Update Mechanism of Author-Topic Model. From a training corpus, we can qualify words-topic distribution and topics-author distribution using author- topic model (ATM). Besides, the author-topic model can be updated with addi- tional documents after training has been completed. This update process is per- formed by Expectation Maximization (EM)-iterating over new corpus until the topics converge. The two models are then merged in proportion to the number of old and new documents. On the other hand, for stationary input (mean that no appearance of new topics in new documents), this process which is equal to the online training of Hoffman

  . If the update process is called with authors that

  already exist in the model, it will resume training on not only new documents for that author but also the previously seen documents.

  Recently, some packages were provided for topic modeling such as topicmodels


  or lda in R, or Gensi in Python. In this study, we choose Gensim for training and updating the author-topic model.

  3 Agent-Based Model for Analyzing Dynamic Social Network Associated with Textual Information (Textual-ABM)

  In this section, we constructed an agent-based model for analyzing dynamic in social network associated with the textual information that we call Textual- ABM (see Fig.

  . We demonstrated dynamic of the social network through the fluctuation of the Textual-ABM in which an agent represent for a network’s node.

  The local interplay among agents transform global environment which includes not only agent’s network structure but also system’s resource, lead to agent’s topic distribution transformation. We present in the following, the steps to build the model and update process for the model in more details.


Fig. 2. Textual-ABM for analyzing dynamic social network using author-topic model


  Dynamic Social Network Analysis Using Author-Topic Model


  3.1 Text Collection and Pre-processing Firstly, we crawl text data from social networking sites or blogs by APIs such as Twitter API, Facebook Graph API and so on. After collecting the data, several pre-processing steps are performed for data cleaning. The textual content obtained is also processed by stemming the words, removing stop-words and numeric symbol. Finally, pre-processed data will be saved into train corpus.

  3.2 Topic Modeling with ATM After obtaining train corpus from text collection and pre-processing, we apply the author-topic model to define topic’s distribution of users. The outputs of ATM contain two matrices: The author-topics distribution matrix θ and the topic-words distribution matrix φ. The topic-terms distribution matrix KxV V

  φ consists of K rows, where the i-th row φ i is the words dis- ∈ R

  ∈ R N xK tribution of topic i. Similarly, the author-topics distributions matrix θ ∈ R K consists of N rows, where the i-th row θ is the topics distribution for i ∈ R author i. A high probability value of θ indicates that author i interested in ij topic j.

  3.3 Textual-ABM Construction Based on the results of topic modeling, we construct a Textual-ABM with the following fundamental elements: Agent. In our model, agents who represent for users in social network, are het- erogeneous with numerous specific characteristics including ID, Name, ID-List (list of agent’s id who have ever interacted), particularly Corpus (collection of texts that an agent used to interact with other agents) and TP-Dis (topic’s prob- ability distribution). Corpus will be cumulated over time through the interactive process. Besides, TP-Dis which illustrates for agent’s interest on topics.

  Agent Interaction. In this study, we take into account directed interaction among agents related to textual information, for instance, retweet or reply on Twitter, sharing statuses on Facebook, collaborating to write papers and so on. Based on the amount of text information of all agents over the interactive process, we not only estimate topic’s distribution of agents at a certain time but also their dynamic over time.

  Agent’s Network. In this study, we consider an agent’s network structure in which a node corresponding to an agent and two kinds of relations among agents: directed interaction relation (R DI ) and major topic relation (R M T P ). R DI is formed since two agents have interacted directly with each other as described in subsubsection Agent interaction, while R M T P appears when two agents are

  52 K. T. Ho et al.

  interested in the same topic with probability greater than threshold p . It can be said that the structural dynamic of network results from agent’s interaction since there is the appearance of new agents, more interactions among agents who have interacted, or new interplay. Moreover, nodes in the network are agents that can change their properties over time, lead to the transformation of similarity or homophily among them.

  Global Environment. Global environment is considered as a space including agents, interaction among agents and resource. Particularly, we emphasize on the system’s resource related to textual information incorporating a system’s corpus and a its generative model in which the former is aggregated from all agent’s corpus while author-topic model (ATM) is used for the later correspondingly.

  3.4 Update Process for Textual-ABM In our model, we consider three principal environmental processes with the pur- pose of updating the Textual-ABM after a certain period of interaction among agents. Firstly, the collection process of new textual data and pre-processing will be performed for generating of a new corpus. Next, the system’s corpus and the existed generative model ATM will be updated with the new corpus. Finally, the Textual-ABM will be updated with the appearance of new agents and agent’s characteristic transformation including ID-List, Corpus and TP-Dis.

  We demonstrated steps for generation of the Textual-ABM and its updating pro- cess at Fig.

   in which the former is formed from step 1 to step 3 while between step 4 and step 6 for the later respectively.

  4 Toy Example

  In this section, we constructed a Textual-ABM for simulating a dynamic network


  For purpose of illustrating dynamic of user’s interest on the small number of top- ics, comment collection and pre-processing are conducted from a political blog


  Firstly, train corpus is collected from 4 pm to 6 pm. A Textual-ABM is con- structed as soon as estimating topic’s distribution of users from train corpus. We considered an agent’s network with two kinds of relationships which include R DI and R with p = 0.1. Furthermore, to reveal dynamic of Textual-ABM, we M T P conducted updated processes with new corpus in next two periods in which the second period lasted from 6 pm to 7 pm and after 7 pm for the third period.

  On one hand, the dynamic of the Textual-ABM is demonstrated through the significant variation of agent’s network structure (see Fig. 2 over three stages. 3


  Dynamic Social Network Analysis Using Author-Topic Model


(a) Period 1 (b) Period 2

  (d) Structural dynamic of some typical (c) Period 3 nodes (0,1,4,7,9) over three periods


Fig. 3. Structural dynamic of agent’s network over three periods: dashed red line with


label in symbol ‘[]’ illustrate R with major topic; solid blue line with label without

DI symbol ‘[]’ reveal for R with the number of interactions (Color figure online)

  The transformation from first to the second period is demonstrated with the appearance of new agent 9, new interactions such as (0, 9) or (4, 9), more interactions between 4 and 7 or between 0 and 1, particularly occurrence R M T P such as topic [2] between 0 and 1. Besides, it is the notable appearance of R M T P between the second stage and the third stage, for instance, (4, 7) are interested in topic [2].

  On other hand, the dynamic of the Textual-ABM is also expressed through agent’s interest transformation since ATM model is updated in the next two periods. We illustrated topic distributions of four representative agents θ through three stages in Table

  We can see that agent ID 0 originally is interested in three

  topics 2, 3 and 1. However, there is a significant variation from the first stage to the second stage, and maintaining this state to the third stage. In contrast, agent

  ID 4 remain interested in first two stages, but there is a drastic change in the third stage. Hence, these fluctuations are the results from the agent’s interaction since the interaction of agent ID 0 to each other is shown mainly in the second period while agent ID 4 in the third period.


  4 K . T . H o et a l.

  Table 1. Topics-Author distribution θ over three periods User ID 0 User ID 1 User ID 4 User ID 7 Topic Prob. Topic Prob. Topic Prob. Topic Prob.

  Period Period Period Period













  3 2 0.542 0.397 0.397 3 0.991 0.231 0.231 1 0.919 0.919 0.003 2 0.968 0.977 0.983 3 0.277 0.412 0.412 0 0.002 0.015 0.015 0

  0.02 0.02 0.491 4 0.008 0.006 0.004 1 0.175 0.187 0.187 4 0.002 0.001 0.001 2 0.02 0.02 0.501 3 0.008 0.006 0.004 4 0.003 0.002 0.002

  1 0.002 0.001 0.001 4

  0.02 0.02 0.003 0.008 0.006 0.004 0.003 0.002 0.002 2 0.002 0.752 0.752 3

  0.02 0.02 0.003 1 0.008 0.006 0.004 Dynamic Social Network Analysis Using Author-Topic Model


  In summary, Textual-ABM can be utilized for analyzing a dynamic social network in which users communicate with each other through text information. The dynamic is not only illustrated by network structure but also fluctuation of agent’s interest.

  5 Independent Cascade Model Based on Homophily (H-IC)

  Information diffusion has been widely studied in networks, with the aim of observing the propagation of information among objects when they are con- nected with each other. There are many diffusion models have been proposed including linear threshold (LT) model, independent cascade (IC) model and so on

  . In IC model, the infected probability is usually based on uniform dis-

  tribution. Nevertheless, in fact, the probability of infection from one object to another depends on similarity or homophily among them, for instance, if two sci- entists research the same field, the probability of information spreading between them will be higher than they are in a different field. Therefore, we proposed IC model on agent’s network that infected probability based on homophily, namely H-IC diffusion model. This diffusion model is presented detail on both static network and dynamic network. We present in the following, the latter steps in more details.

  5.1 Independent Cascade Model (IC) We assume a network G = (V, E, P ), where P : V × V → [0, 1] is probability function. P (u, v) is the probability of node v infected from node u. The diffusion process occurs in discrete time steps t. If a node adopts a new behavior or idea, it becomes active, otherwise it is inactive. An inactive node has the ability to become active. The set of active nodes at time t is considered as A t . new

  Under the IC model, at each time step t where A is the set of the newly t − new


  infects the inactive neighbors v activated nodes at time t − 1, each u ∈ A t − out

  1 (u) with a probability P .

  ∈ η (u,v)

  5.2 Similarity or Homophily Measure Between Two Agents Homophily is the tendency of people to associate more with those who are sim- ilar to them on some dimensions than with those who are not. In this study, we measure similarity or homophily between two agents based on topic’s prob- ability distribution. If we consider a probability distribution as a vector, we can choose some distances measures related to the vector distance such as Euclidean distance, Cosine Similarity, Jaccard Coefficient, etc. However, it is better if we choose distances measures related to the probability distribution such as Kull- backLeibler Divergence, Jensen-Shannon divergence, Hellinger distance, etc. In this study, we choose Hellinger distance to measure similarity or homophily. Let two discrete probability distributions, P = (p , p , ..., p k ), Q = (q , q , ..., q k )





56 K. T. Ho et al.

  Hellinger distance: k

  1 √ √


  d p q H (P, Q) = √ ( i i )

  2 i


  Similarity or homophily: H (P, Q) Homo(P, Q) = 1 − d

  Algorithm 1. R-IC diffusion on static agent’s network

  Require: A Textual-ABM (include agent’s network G=(V, E)), A : seed set 1: procedure R-IC-Static-Network(G, A ) all new 2: t = 0, A = A , A = A 3: while activation occur do all 4: t:=t+1; A t = ∅ new 5: for u ∈ A do t 6: Calculate A all all t t t (u) = {v ∈ out(u), p <= q}; p, q ∼ U (0, 1) 7: A = A A (u) all all all new all A = A A t ; A = A t all

  ⊲ 8: return A Output

  5.3 Random Independent Cascade Model on Static Agent’s Network

  In this section, random independent cascade diffusion model is presented on a static network, namely R-IC diffusion model. This model plays a role as a baseline model for comparing performance with H-IC diffusion model that we will propose at Sect.

  Firstly, diffusion starts with a set of active agents.

  We consider new active agents which can active their inactive neighbors with a probability randomly. After the initiation step, in each round, the propagation continues until no more activations can occur (see Algorithm .

  5.4 H-IC Diffusion Model on Static Agent’s Network In this section, H-IC diffusion model is proposed on a static network where network structure fixed through the diffusion process. Firstly, diffusion starts with a seed set and new active agents which can active their inactive neighbors with a probability based on homophily among them. Obviously, if homophily between an active agent u and an inactive agent v is higher then the diffusion probability is greater. The spreading process continues until no more activations can occur (see Algorithm .

  Dynamic Social Network Analysis Using Author-Topic Model


  Algorithm 2. H-IC diffusion on static agent’s network

  Require: A Textual-ABM (include agent’s network G=(V, E)), A : seed set 1: procedure H-IC-Static-Network(G, A ) 2: t

  = 0, A all = A , A new = A 3: while activation occur do 4: t:=t+1; A all t = ∅ 5: for u ∈ A new do 6: Calculate A t (u) = {v ∈ out(u), p <= homo(u, v)}; p ∼ U (0, 1)

  7: A all t = A all t A t (u) A all = A all A all t ; A new = A all t 8: return A all ⊲ Output

  5.5 H-IC Diffusion on Dynamic Agent’s Network Although IC model on dynamic network has studied by

   ]. However, the

  dynamic of a network is only presented through structure variation, while infected probability from an active node to inactive node is always fixed. There- fore, in this study, we proposed H-IC diffusion model on a dynamic agent’s network in which not only include the transformation of network’s structure but also the variation of agent’s topics distribution. It can be said that spreading probability among them can change over time since homophily among agents transform.

  The dissemination initiates with a Textual-ABM which include an agent’s network G, a seed set contains active agents A and an updated set U including steps k that at those steps the Textual-ABM will be updated. New active agents which can active their inactive neighbors with a probability based on homophily among them. During diffusion process, at steps, k ∈ U, agent’s network G will be updated once since Textual-ABM is updated. The spreading process will run with n steps (n > max{U}) (see Algorithm .

  Algorithm 3. H-IC Diffusion on dynamic agent’s network

  Require: A Textual-ABM (include agent’s network G = (V, E)); A : seed set

Require: U = {k}, at step k G is updated; n: number steps of diffusion (n > max{U })

1: procedure H-IC-DYNAMIC-NETWORK(Textual-ABM, A , U

  ) 2: t = 0, A all = A , A new = A

  3: while t < n do 4: t:=t+1; A all t = ∅ 5: if t ∈ U then: 6: Update Textual-ABM ; A


= A all 7: for u ∈ A new do 8: Calculate A t (u) = {v ∈ out(u), p <= homo(u, v)}; p ∼ U (0, 1)

  9: A all t = A all t A t (u) A all = A all A all t ; A new = A all t 10: return A all ⊲ Output

  58 K. T. Ho et al.

  6 Experiments

  6.1 Dataset The proposed H-IC models were evaluated on scientific collaboration network where nodes are scientists and links are co-authorships since cooperating in common articles. We have targeted authors who have participated in Neural Information Processing Systems Conference (NIPS) from 2000 to 2012. Textual data is collected and archived annually. The dataset contains 1740 papers which are contributed by 2479 scientists.


(a) Number topics in range [10, 200] (b) Number topics in range [50, 90]

Fig. 4. Log-likelihood

  6.2 Setup Firstly, we gathered textual data between 2000 to 2008 for train corpus and defined the author’s topic distribution using ATM. We chose the number of topics based on the Harmonic mean of Log-Likelihood (HLK)

  . Firstly, we

  calculated the Harmonic mean of Log-Likelihood with the number of topics in the range [10, 200] with sequence 10. We realized that the best number of topics are in the range of [50, 90] of a maximum value of HLK (see Fig.


  . Therefore, we ran again with the number of the topic in [50, 90] with sequence 1 and obtained the best is 67 (see Fig.


  ). Based on these results, a Textual-ABM is constructed. Notice that we only consider the diffusion on agent’s network with relation R DI , that means agent A can active agent B if A have relation R DI with B.

  For experiments about H-IC diffusion on static agent’s network, firstly, we constructed four Textual-ABMs. Next, the first experiment is started on agent’s network which is formed as soon as the first Textual-ABMs is updated with new corpus in 2009. Similarity, the remaining three experiments are implemented on agent’s networks which are generated after the Textual-ABMs are updated two times, three times and four times respectively. For each static network corre- sponding each year from 2009 to 2012, we just implement the diffusion procedure

  Dynamic Social Network Analysis Using Author-Topic Model


  in the largest isolated community H. A million random samples are used for seed set and a million times of diffusion process is executed for each seed set. Besides, we also conducted random independent cascade model (R-IC) as a benchmark to compare the performance of H-IC.

  For experiments to simulate H-IC diffusion on dynamic agent’s network, the diffusion process starts as soon as the Textual-ABM is formed. During diffusion process, the Textual-ABM is updated. We implemented four experiments in which the first one is the diffusion on agent’s network without dynamic. The last investigations are that after every five steps, ten steps and fifteen steps of diffusion agent’s network will fluctuate once since the Textual-ABM is updated.

  6.3 Model Evaluation To evaluate the performance of diffusion models, we use the number of active nodes or the active percentage which are standard metrics in information diffu- sion field

   ]. In this research, we utilized the active proportion to evaluate

  the performance of spreading models on the static network and the number of active nodes for diffusion models on a dynamic network.

  We compare the performance of proposed H-IC diffusion model with baseline model. This baseline model is random independent cascade model (R-IC) that we presented at Sect.


  6.4 Results H-IC Diffusion on Static Agent’s Network. Diffusion results on static agent’s network are shown in Fig.

  We can see that active percentage of H-

  IC diffusion been always greater than R-IC diffusion in all four years. In 2009, the active proportion reaches approximately 0.08% for R-IC diffusion while H-

  IC diffusion attains around 0.2%. Moreover, the active rate of H-IC diffusion is three times, two points five times and twice greater than R-IC in 2010, 2011 and 2012 respectively. Therefore, based on the active percentage we can conclude that H-IC diffusion outperform comparison with R-IC diffusion model.

  Furthermore, results demonstrated that the diffusion corresponds to the dynamic of networks overtime since next year’s rate of both H-IC and R-IC is always higher than the previous year. The active proportion of R-IC obtains about 0.08% in 2009 and increases up to 0.117, 0.25 and 0.4% in next three years correspondingly. Besides, the percentage of active agents of H-IC in 2010 is approximately 0.4% which is twice as large as in 2009. Moreover, there is a significant increase in the active percentage of H-IC comparison with 2010 since it reaches 0.622% in 2011 and its peak at about 0.7% in 2012. It can be said that the structural transformation of agent’s network over the year lead to the fluctuation of the active rate of the diffusion process.

  H-IC Diffusion on Dynamic Agent’s Network. Results are shown in Fig.


  which illustrate the spreading process without dynamic of network quickly stop

60 K. T. Ho et al.


Fig. 5. Diffusion on static network

Fig. 6. Diffusion on dynamic network

  at the 8th step, while there is a significant fluctuation in the active number if agent’s network has transformation after some steps of diffusion. We took into account the active number and steady state of diffusion for both static and dynamic agent’s networks. Firstly, the propagation process reaches approx- imately 227 active agents which are remained from the 8th step onwards in a static network. On another hand, we can see that the active number of diffusion process on three dynamic agent’s networks obtain around 450. Particularly, the

  Dynamic Social Network Analysis Using Author-Topic Model


  spreading process sometimes falls into a steady state for instance between 8th and 10th step or 16th and the 20th step for agent’s network that has transforma- tion after each ten step of diffusion. However, this status is broken since agent’s networks are dynamic. In short, it can be concluded that the diffusion will reach and remain steady state on a static network while it can be transformed if agent’s network fluctuates during the diffusion process.

7 Conclusion

  In this paper, we constructed an agent-based model for analyzing dynamic social network associated textual information using author-topic model, namely Textual-ABM. The benefit of this model is to can be utilized to observe dynamic of a social network which includes not only network structure but also node’s properties. Moreover, we proposed an independent cascade model based on homophily, namely H-IC. The probability of infection between two agents is measured by homophily or similarity based on topic’s distribution. H-IC is illus- trated on both static and dynamic agent’s network. Results from experiments demonstrated that the effectiveness of H-IC on the static network outperforms comparison with R-IC. In addition, experiments also demonstrated the fluctua- tion of the active proportion of the spreading in dynamic agent’s network instead of obtaining and remaining a steady state in a static network.



1. Breiger, R.L., Carley, K.M., Pattison, P.: Dynamic Social Network Modeling and

Analysis: Workshop Summary and Papers. National Academies Press, Washington,

D.C. (2003)


2. Buntine, W.: Estimating likelihoods for topic models. In: Zhou, Z.-H., Washio,

T. (eds.) ACML 2009. LNCS (LNAI), vol. 5828, pp. 51–64. Springer, Heidelberg (2009).


3. Carley, K.M., Martin, M.K., Hirshman, B.R.: The etiology of social change. Topics

Cogn. Sci. 1(4), 621–650 (2009)


4. Gayraud, N.T., Pitoura, E., Tsaparas, P.: Diffusion maximization in evolving social

networks. In: Conference on Online Social Networks, pp. 125–135 (2015)


5. Girvan, M., Newman, M.E.J.: Community structure in social and biological net-

works. In: Proceedings of the National Academy of Sciences of the United States of America, pp. 7821–7826, June 2002


6. Grandjean, M.: A social network analysis of Twitter: mapping the digital human-

ities community. Cogent Arts Humanit. 3(1), 1171458 (2016)


7. Hoffman, M., Blei, D.M., Wang, C., Paisley, J.: Stochastic variational inference.

[cs, stat], June 2012

  8. Holme, P., Saramki, J.: Temporal networks. Phys. Rep. 519, 97–125 (2012)


9. Legendi, R.O., Gulys, L.: Agent-based dynamic network models: validation on

empirical data. In: Advances in Social Simulation, pp. 49–60 (2014)

10. Blei, D.M., Ng, A.Y., Jordan, M.: Latent Dirichlet allocation. J. Mach. Learn. Res.

  3, 601–608 (2001)

  62 K. T. Ho et al.


11. Newman, M.E.J.: Fast algorithm for detecting community structure in networks.

  Phys. Rev. E 69, 066133 (2004)


12. Niazi, M., Hussain, A.: Agent-based computing from multi-agent systems to agent-

based models: a visual survey. Scientometrics 89(2), 479–499 (2011)


13. Otte, E., Rousseau, R.: Social network analysis: a powerful strategy, also for the

information sciences. J. Inf. Sci. 28(6), 441–453 (2002)


14. Rosen-Zvi, M., Griffiths, T., Steyvers, M., Smyth, P.: The author-topic model for

authors and documents. [cs, stat], July 2012


15. Shakarian, P., Bhatnagar, A., Aleali, A., Shaabani, E., Guo, R.: Diffusion in Social

Networks. Springer, New York (2015). Google-Books-ID: zouQCgAAQBAJ

dynamic social networks. In: International Conference on Data Mining, pp. 1313– 1318, December 2013


Concept of Temporal Data Retrieval Undefined

Value Management

  Michal Kvet (&) and Karol Matiasko


Faculty of Management Science and Informatics, University of Žilina,

Univerzitná 8215/1, 010 26 Žilina, Slovak Republic

Abstract. Data validity, reliability and correctness of the results form the core

of information system. Each data tuple can be bordered by the time definition.

Multiple temporal approaches have been developed for different conditions and

environment properties. One of the most significant data aspect is just the time.

Solutions, in which data were updated in a strictly expected way, granularity and

frequency are over. Therefore, we propose an attribute-oriented temporal model

with reflection on data grouping technologies. Many times, data obtained from

the sensors or measuring equipment are not relevant, reliable or are provided too

late. For these reasons, undefined values must be stored in the database in the

form of time itself or attribute expressing the state of the object. This paper deals

with undefined values compromising performance and data access methods. We

also propose definition and management using database objects and pointers. In

the performance evaluation section, multiple indexes are compared to declare

performance of our proposed solution in comparison with existing methods.

  Keywords: Temporal characteristics Attribute-oriented approach Index NULL value DataObjTimeSolution UndefTypeValue

1 Introduction

  Data validity, reliability and correctness of the results create the main part of each information system. Development principles, characteristics, approaches and methods evolve over the time. One of the most significant data management characteristics is just the time frame. It is not only the area of time management in the spectrum of validity but also transaction definition, position or conditions should be evaluated. Significant evolution can be perceived from the 80s of the 20 th century when the technology became simpler and easier to manage. Moreover, the strong impact had also fallen in hardware prices and the ability to manage larger data volumes. Even in this phase, conventional approaches were mostly used by covering creation and standard- ization of the existing paradigm, which is based on storing only current valid data. Thus, each Update statement automatically replaces the original value. Temporality evolution has started with log files managed by the flashback technology followed by the architectural solution – object and attribute-oriented approach. There were several attempts to create a temporal standard, unfortunately, it has not been approved yet [


  64 M. Kvet and K. Matiasko

  Nowadays, research streams have been extended using multiple ways. The real-time database uses real-time processing to handle workloads, whose state is constantly changing. It manages dynamic attribute value changes. The conventional database cannot be transformed into real-time solution, because of the inconsistencies between the real world objects and the data that represent them are too severe for

  Another system is formed based on future databases, which

  simple modifications [ main target is the management of future states, to evaluate future conditions and create plans for responses and solutions. In the recent past, temporal database approach has been extended by the locality position forming spatiotemporal database approach, for tracking moving objects determining the single position at defined time point. Com- munication is usually provided by ad-hoc networks with regards to reliability. In this sphere, also attribute-oriented approach shows its complexity and robustness – com- munication is strictly bordered by the time, afterwards, communication is not possible due to connection loss in quality and speed [

  The temporal approach can be also

  managed using activity modelling ].

  In this paper, we deal with temporal modelling characterized with undefinition of the attribute values, Sensorial data constitute a substantial part of the intelligent systems and should be managed effectively with regards on attribute granularity and frequency of changes. The significant aspect is just undefined value and state management.

  2 Own Hybrid Architectural Solution

  The area of the temporal database is also part of our interest. We, therefore, propose another solution, which can be considered as attribute oriented, however, it does not use it in the pure level. In principles, sensorial temporal data provided by intelligent systems are characterized by various granularity and frequency of changes. Therefore, the attribute-oriented approach is far more suitable in comparison with object-oriented temporal approach [

   ]. Thus, each change of the attribute delimits execution of the

  new Insert statement to the second (temporal) layer. It just may cause problems and performance reductions. The principle covers the individual modules, that provide sensory data or data, which values evolve over the time and should be monitored regardless the structure. These modules can preprocess and evaluate the input data and provide them synchronously. Though, if the data structure is structured by interval dependencies, it would cause multiple Insert statements with reflection to the historical data transformation. On the other hand, there should always be a possibility to deal with pure data with no synchronization. For these purposes, a new approach is intro- duced. It extends the principles of the second layer of the attribute-oriented approach by group definition. To eliminate the impacts of synchronization module failure or processing delays, the temporal system can also communicate with the data provider directly. Thus, three situations can occur:

  • Pure data are always processed by direct connection of the data provider to the system.
  • Data are always sent to synchronization and preprocessing module, thus, there is no direct connection to the system. It is used to eliminate huge workload processing
and evaluating of the system by transferring module to another segment or node of the network. Thanks to that, input data of the application are compressed and deprived of incorrect data (e.g. caused by communication cable problems or packet loss).

  • Universal solution is based on two line communication possibilities. Obviously, data are sent to synchronization module for processing and evaluating, but if there is significant delays or data should be provided, but have not been available yet, the temporal manager can request data by accessing data provider. In this case, how- ever, a strict emphasis on timestamps must be done to ensure, that late data from the module would not influence directly obtained data.

  Synchronization module provides and ensures several activities, which can be processed either sequentially or in parallel. It concerns on input data cleaning, pre- processing and data buffer caching, which are then sent in batches. Data can be optionally encrypted. A special challenge is just the data evaluation in terms of reli- ability, consistency with regards to undefined values. Many systems are sensitive to the quality, form and time of the required data. Thus, if they are not possible to obtain data, particular information must be stored in the database. Accordingly, if the system evaluated, that the environment properties and characteristics changed, but we have no new object states, previous states must be denoted as historical with no actual data. Thus, undefined values are stored, which can cause significant performance degrada- tion during the data retrieval operations.

  In the previously defined hybrid environment, several invalidity situations can occur. For the clarity and explanation simplicity, let’s assume, that the provider data are not grouped. If the data provider is connected directly to application system and data are not provided or the system evaluates, that they are not correct, an undefined value is stored for the particular attribute or group of attributes, if the provider can offer multiple data types. In this case, it is necessary to highlight, that no matter the number of data provided by the data node, it is important, that such data are not synchronized (grouped). The much more complicated situation arises, however, when synchroniza- tion module layer is used. Therefore, we introduce the term data value (DV) covering the following three situations: • Data attribute value is unknown and not grouped.

  Concept of Temporal Data Retrieval Undefined Value Management


  • Data attribute value is part of the group and the whole is denoted as unknown. In this case, there is the very similar solution to the data attribute itself.
  • Data attribute value is unknown and grouped, but the rest part of the group is still correct and valid. If only group identifier is stored in the database, the whole group must be denoted as invalid reflecting unknown data, which is not very useful, because even valid data must be replaced. Our proposed solution is based on temporal group definition, thus the participation of the attribute is also time delimited. Again, data interval overlapping for the group is prohibited, the attribute cannot be part of more than one group at the same time point or interval. Simpli- fication of the issue provides the solution, which can form a group consisting only of one attribute. Such solution techniques do not manage data provider and pure data at all, but they are virtually wrapped as a single member group.

66 M. Kvet and K. Matiasko

  An individual state is formed by several attributes. During the event of state update, some attributes can be changed into definite values, some portions are not changed and the rest part is undefined. To ensure validity and consistency of the model, original group is divided into two parts (Group_part1 and Group_part2). The first group (Group_part1) contains attributes, which values are correct or not changed, the second group (Group_part2) covers invalid attributes. If the invalid attribute value will become valid in the future, after the particular update, divided groups will be merged together. The group part identification is based on the naming convention – the basis is the original name of the class followed by the fragment identifier, therefore there is no problem to find all group division parts.

  Physical representation (shown in Fig.

  is based on data model with attribute

  (Attribute entity) and group identification (Group entity) covering assignment evolution over the time. It is modelled using ISA hierarchy highlighting time (1:N relationship type). In principles, also the single attribute can be considered as single item set. Group can be divided into parts (Part entity) over the time, which is physically managed by time association and assignments. These data are stored in Association entity grouping fragments (parts). It is important to emphasize, that not an only group but also group parts themselves can be divided. For these purposes, following notation is used (as- sumption – Group1 is defined, which is divided into two parts – Part1 and Part2. As the result of the transaction Update statement, defined Part1 will be divided into two groups – Part1 and Part2):

  Group1 => Group_part1, Group_part2 => Group_part1_part1, Group_part1_part2, Group_part2. Thus, in the end, we have 3 fragments. Whereas Group_part1_part2, Group_part2 are invalid, they can be grouped together by postprocessing. The final result is: Group_part1_part1 (VALID), Group_part1_part2 (INVALID), Group_part2 (INVALID) => Group_part1 (VALID), Group_part2 (INVALID).


Fig. 1. Physical representation


deals with existing data access methods followed by Sect.

   proposing our own solution.

  67 Concept of Temporal Data Retrieval Undefined Value Management

3 Data Access – Paths, Index Structures, Existing Solutions

  Each Data Manipulation Language statement is covered by the transaction. After reaching Commit operation, no data can be lost at all, although the data changes have been physically made only in the memory buffer cache structure. Two-phase Commit protocol always ensures that transaction log is created physically, so UNDO and REDO data are generated and stored physically in the storage sector. On the other hand, data to be processed must be loaded into memory before the processing itself. Database Writer process then writes particular data physically, however, it is not done immediately due to workload optimization and lazy loading principles. Physical data representation is reflected by the data files located in the storage covered by security options. To ensure the independence of database and instance, the interim layer is defined – tablespace.

  Enforcing and improving performance can be done by various techniques, the main aspect defines an index, which is optional data structure associated with a table, which allows the optimizer to easily locate particular data object tuple. As a consequence, the strong data stream can be obtained automatically without the necessity to access physical data file at all. Thanks to that, the number of I/O operations is reduced, whereas data are distributed in blocks randomly based on current free size. The absence or presence of an index does not require a change in the wording of any SQL statement. An index is a fast access path to a single row of data. It affects only the speed of execution. Given data value that has been indexed, the index points directly to the location of the rows containing that value. Database management system automatically maintains the created indexes – changes (Insert, Delete, Update) are automatically reflected in index structures. However, the presence of many indexes on a table degrades the performance because the database must also update the indexes.

  There are several index structures, which can be maintained and defined in the relational databases. In this paper, we will deal with solutions delimiting undefined values in the temporal environment. The general structure forms currently B+tree index, which has been developed for conventional databases, however, they are very suitable and effective for temporal structures, too. The main characteristic of them is just effectivity and performance, which is not significantly influenced by the frequent data changes, whereas temporal environment mainly enhances Update and Select statements. It consists of a balanced tree, in which each path from the root to the leaf has the same length, which can require rebalancing during new node adding, removing or position changes. In this structure, we distinguish three types of nodes - root, internal node and leaf node. Operations, principles, code and rebalancing methods can be found in

   ]. Several improvements and optimization can be done to improve performance.

  In the past, undefined values were mostly modelled using NULL values, which, as we will describe later, is not quite correct. The problem of standard B+tree index covers the fact, that is cannot store and manage NULL values – such values cannot be com- pared using relational operations, cannot be sorted directly. As a consequence, if the attribute of the query contains (or even can contain) NULL values, the particular index cannot be used, resulting in using the Table Access Full method. The aim is, therefore, to eliminate, respectively replace NULL values. How to do it?

68 M. Kvet and K. Matiasko

  First of all, it is necessary to mention, that NULL value does not have special denotation and meaning. From such value, in principle, no information about the real value can be obtained. Data validity can be modelled using various characteristics. In the past, several attempts to standardize data type period have been done, however, no one resulted in certification. Thus, temporal data interval definition must be done explicitly and managed manually. Generally, therefore, the interval is framed by begin (BD) and end (ED) value based on the granularity. BD always expresses the exact value delimited by arrival time or time of processing. In case of delays, a particular value can be bounded by the physical time. Limiting the validity of the right site by ED can be insertion. Therefore, undefined value modelled by NULL definition can be considered as a suitable solution. However, if object level temporal characteristics are used, it cannot be used at all from the definition. The original primary key is extended by the validity definition and no part of the primary key cannot express undefined value. The important factor is just carried by the validity interval itself. Although many times, exact time limitation cannot be done, it can be confidently said, that it reflects the future value, simply, such time point has not occurred, yet, which brings major significance in comparison with the absolute value modelled by NULL value notation. Moreover, notice, that NULL value cannot be compared using mathematical operations at all. Therefore, UndefinedTimeSpectrum notation has been introduced.

  In this paper, more effective solutions to cover temporal NULL values character- istics have been proposed by us, which are briefly introduced in the following part of the paper, whereas they are also compared with our proposed solution technique.

  The first approach is a function-based index as a special type of B+tree access structure. It does not manage direct attribute value, but it based on computing the output values of an expression. By this approach, the physical NULL value can be stored in the database, however, during data retrieval process, NULL values are replaced with undefinition sign. Therefore, the function-based index can be very useful, if actual statistics reflecting model can be provided. If so, Index Scan method can be used instead of accessing all data blocks. Usually, the amount of output data is rela- tively lower in comparison with all objects and evolution over the time, thus, a function-based approach using B+tree index can provide an efficient solution. Function-based index stores precomputed function values directly with the pointers (ROWID) to data themselves. On the other hand, the main disadvantage is, that opti- mizer can use a function-based index only for cost-based optimization, not for rule- based optimization . Other disadvantages are based on indeterminate length manage- ment, thus data types varchar2, raw, long raw and some other PL/SQL data types cannot be processed at all. Naturally, the function-based index cannot invoke aggregate functions. In a temporal database, it can evaluate validity interval definition, time length and position of the validity, but can be also used to replace the NULL value of the undefined attribute value. Thus, the solution is universal for any data type or structure. Moreover, the result of the function can be stored in the result_cache memory to simplify its evaluation.

  Another approach is characterized by virtual column, which value is provided by the function. Afterwards, such virtual column is indexed. Virtual column means, that is

  69 Concept of Temporal Data Retrieval Undefined Value Management

  not physically stored in the database, but index approach works like it would be there. The additional attribute value is dynamically computed on the fly.

  Intelligent systems and complex data management require exact state definition with no opportunity to mix undefined states with states, which validity is delimited by the future time point. Therefore, researcher offered new notation for dealing with undefined time limitation managed by MaxValueTime notation. Such undefined values are physically replaced by the notation. Thanks to that, it is easy to locate actual time non-limited states. This approach requires the same data type as the associated attribute data type, which can sometimes require too much space.

  Some other approaches dealing with undefined values can be found in [

  4 Own Solution – DataObjTime, UndefTypeValue Database Object

  To point out the management of undefined values, we propose own solution (DataObjTime), which is based on new database object definition in the data dictionary. At the first glance, it might seem that the solution is similar to previous ones. It is based on processing, evaluating and eliminating NULL values, but in our case, we do not use the function result nor virtual column or even MaxValueTime notation, but highlight the solution using object and pointer to it, not physical value. The solution can be con- sidered as a constant object, which is automatically loaded into memory after mounting and opening database. Thanks to that, no evaluation, loading and invalidity management is necessary. Such solution is obviously defined for each data type separately, however, if transformable and compatible data types with each other were used, it would be technically possible to use the common object. Be aware, such solution would not be optimal in terms of implicit conversion necessity. Thus, better performance would provide a solution using explicit conversion functions, however, the ideal solution creates an individual notation for each used data type, which associated attribute can contain undefined values. In the memory, it has very low resource requirements, each object is namely characterized by only one bit (1 = DataObjTime is active;

  = DataObjTime is deactivated). Physical structure in the database deals with temporal evolution, it is also based on column level temporal system delimiting the activity and usability of proposed notation. First of all, definition object must be created. For these purposes, we offer a script to create definition objects, consisting of one or two attributes provided in the package (UNDEF_OBJECT package). The first parameter characterizes the data type in the string format, to which reference should be created and associated. The second parameter is optional, binary format, which can get two values (1 = DataObjTime is active; 0 = DataObjTime is deactivated). By default, the solution is enabled automatically. Executing such script also ensures, that information about its activation and creation is stored in particular temporal table in data dictionary:

  Undef_object.create (p_data type varchar [, p_status integer])

  70 M. Kvet and K. Matiasko

  The previously defined concept is manual – definition data objects are created explicitly by the user. The second proposed solution is more complex and user-friendly. Attribute considered to be temporal must be registered in column level approach. Thanks to that, the list of data types can be listed, for each of them, temporal undefinition object can be created. It is provided by the function Create_all_objs function, which result expresses the number of created notations. If the notation is already created, it is ignored and existing one is used. Using overloading, the Cre- ate_all_objs function can be called without any parameters, or just one can be used (whether the evolution management and monitor are enabled or not. If an attribute with new data type is loaded, it delimits, whether definition object is created automatically, or not):

  Undef_object.create_all_objs (p_level integer)

  Object references are managed by the pointers, which value size is 1 byte. Thus, it also lowers disc space requirements (e.g. Date requires 7 bytes, Timestamp even 11 bytes). MaxValueTime notation in comparison with our solution reflects the same size as an associated data type.

  5 Performance

  Our experiments and evaluations were performed using defined example table – employee with the following structure:

  • Id – the personal number of employee,
  • BD, ED – time period of the row validity,
  • Name, Surname,
  • Dept_id - employee department association, • Salary.

  50 departments were used, each consisting of 1000 employees, each of them was delimited by 20 different salaries over the time. Thus, a total number of rows was one million. No primary key was defined, because of the environment properties and our direct opportunity for explicit index definition.

  Experiment results were provided using Oracle Database 11g Enterprise Edition Release - 64 bit Production; PL/SQL Release – Production. Parameters of used computer are: processor: Intel Xeon E5620; 2,4 GHz (8 cores), operation memory: 16 GB and HDD: 500 GB.

  To point out management of undefined values over the time, it is necessary to evaluate our proposed solution with existing approach to declare performance as well as define limitations. Whereas temporal characteristic requires each state to be defined by no more than one row, our defined environment limits the number of actual states to 50 000. In the following experiments, various number of actual states is used – 5%; 2%; 1%; 0,5; 0,2% and 0,1%:

  71 Concept of Temporal Data Retrieval Undefined Value Management

  In the first phase, comparison of undefined time value denoted by a NULL value in comparison with DateObjTime database object solution (our developed model) is used. B+tree index based on attributes ED, BD and ID is created (in such defined order) – the reason and comparison of index types can be found in [

  The select clause of the

  statement consists of only ID attribute, thus, all data can be obtained using an index

   shows the

  with no necessity for accessing data files in the physical storage. Figure experiment results. As we can see, if NULL values are used, there is no other way, so Table Access Full (TAF) method must be used to avoid NULLs. If the undefined value is modelled by our DateObjTime solution, all values are indexed, so Index Range Scan with significant performance improvement can be used. Total costs and pro- cessing time reflect significant performance growth, too. If all data have actual non-limited value, 86, 67% of costs are eliminated, which reflects 86, 96% of pro- cessing time. With the reduction of the number of undefined values, the difference is even more strict – 99, 56% of costs and 86, 96%. Processing time does not depend on a number of actual data ratio. The reason is based on the necessity of index loading into memory, which reflects the same time. On the other hand, total costs cover not only memory but also other server resources, are eliminated with data dimension down tendency. The special category covers Table Access Full method, which must load and evaluate all data blocks of the table covered by High Water Mark (which does not transfer its value to lower segment).


Fig. 2. Performance results – NULL, DateObjTime

  The second part deals with the extension of the Select clause. In this section, more attributes that are part of the index are required, therefore data must be accessed by Rowids if index method is used. Again, NULL value performance is really worse. It does not express significantly different results in comparison with the previous experiment. It always reflects all data blocks, which are loaded into memory. Data processing and transferring into the result set from memory is negligible. DateObjTime solution offers a wide range of performance improvement possibilities. Also, in this case, Index Range Scan method can be used, which must be, however, followed by loading particular data block to provide data, which are not part of the index. This method is called Table Access by Index Rowid (TAIR). It reflects the slowdown from 8% up to 25% in our environment. Average value of the added costs is 13,35%. Figure shows the results.

72 M. Kvet and K. Matiasko


Fig. 3. Costs– NULL, DateObjTime solution

  The third part deals with comparison with current sophisticated solutions for dealing with replacing NULL values. In this stage, four sophisticated solutions (func- tion based, an index based on the virtual column, existing notation and new object approach) are compared based on costs and processing time. For the evaluation, we will use the same environment as previously defined, actual data ratio used is 1%. Thus, it reflects the value 1840 of costs, when NULL values are directly processed, 104 for costs of our proposed DateObjTime object solution. When dealing with processing time, it reflects 24 s for an original solution and 3 s for our own approach. For the clarity, we introduce these labels:

  • Model_1 – there is no special solution for dealing with NULL values (reference model).
  • Model_2 – NULL values of the time attributes are replaced using the function-based index.
  • Model_3 – virtual column is used replacing time NULL values. Such column is consequently indexed.
  • Model_4 – existing MaxValueTime notation is used. This solution characterizes NULL values with maximal date value – physical representation.
  • Model_5 – our proposed solution, which uses DateObjTime approach using a pointer to the database object. The undefined value is therefore modelled logically.


   shows the performance (costs) of the proposed models. When dealing with

  reference model compared to existing solutions, the worst performance provides a function-based index, however, also function based solution provides significant per- formance improvements – 91,6% for costs. The reason is based on function value

  73 Concept of Temporal Data Retrieval Undefined Value Management

  evaluation. The first problem can arise with the necessity to recompile function after some change inside. The second parameter highlights the physical memory structures and representation. Function results are stored in the result cache memory structure, whereas index and data themselves are stored in database buffer cache, which can have another structure with different block size or can be stored in another segment. Model_3 is based on the virtual column, it provides improvements in rate 93, 16% for costs (in comparison with reference Model_1). In this solution, when data are loaded into memory, virtual column value is evaluated for the data themselves as well as for defined index. It is stored in the shared buffer cache memory structure and accessed improvements using 18,18% rate. Model_4 uses optimized solution using MaxValue- Time notation. It avoids function definition but uses automated time NULL values management mostly defined by trigger events. In comparison with the reference model, it provides performance improvements – 93, 70% for costs. In comparison with the best solution provided (Model_3), it reflects improvements – 7, 94%. the reason is based on no necessity to evaluate any value, no function result – replaced values are already in the database, physically stored. Our proposed DateObjTime solution is characterized by Model_5 . Physical representation is improved, the logical view provides a value rep- resenting future time, which has not occurred yet, however physical expression is delimited by a pointer to the database object in the data dictionary. It lowers the storage requirements to one-seventh for a date, one-eleventh for timestamp (pointer solution requires only 1 byte for the processing. Moreover it is possible to group pointers to eliminate disc storage requirements more deeply). Comparison of the Model_4 with Model_5 is characterized by 10,34% improvements in costs.

6 Conclusion

  Temporal enhancement of the database systems brings the opportunity to store, manage and evaluate data during the whole lifecycle of the objects. Data management, as well as performance, is an inevitable part of the solution. In the first part of the paper, the introduction of existing temporal approach is referenced with regards to data structure definition. The column-level model uses attribute granularity and manages each attri- bute separately. It provides an efficient solution for sensor-based architectures but does not guarantee sufficient power, if some attributes are grouped or synchronized. In that case, each group would be divided into single attribute instances. Therefore we propose our own hybrid solution, which combines object and column level, temporal model.

  The second part of the paper deals with undefined state management, which can be effectively managed only in column level or hybrid solution because object level must associate the whole state as invalid regardless the number of valid attribute values. For dealing with undefined states and actually unlimited validity, existing solutions are described with regards to access paths and index structures. Existing solutions are mostly characterized by index approach by removing the impact of NULL values. Another solution is based on notation definition replacing NULL values - MaxValue- Time . To participate the performance, we propose our own solution - DataObjTime, which is based on database object stored in the data dictionary, to which pointers are

74 M. Kvet and K. Matiasko

  used. Thanks to that, disc requirements are lowered and performance is improved. Comparison of proposed and existing solutions is provided in the experiment section, where multiple solutions are compared. Our defined DataObjTime solution provides the best performance in both categories – costs and processing time. In comparison with existing MaxValueTime notation, improvements in costs are more than 10%, with other models, results are even more significant (more than 17% in comparison with index based on the virtual column, more than 32% for function based index). In case of comparison with pure B+tree index, costs can be lowered up to 94%, by removing NULL values identified by time, which cannot be compared using relational rules and

  In the future, we would like to extend the defined solution to a distributed environment.

  Acknowledgement. This publication is the result of the project implementation:

Centre of excellence for systems and services of intelligent transport , ITMS 26220120028

supported by the Research & Development Operational Programme funded by the ERDF and

Centre of excellence for systems and services of intelligent transport II.,

  ITMS 26220120050 supported by the Research & Development Operational Programme funded by the ERDF.

This paper is also the result of the project implementation Center of translational medicine,


ITMS 26220220021 supported by the Research & Development Operational Programme funded

by the ERDF.



1. Blef, J., Fernando, L., Foster, K., Abhilash, G., Sudarsan, S.: Spatio-temporal reasoning and

decision support for smart energy systems. In: 2016 IEEE 21st International Conference on Emerging Technologies and Factory Automation (ETFA) (2016)


2. Doroudian, M., Arastouie, N., Talebi, M., Ghanbarian, A.: Multilayered database intrusion

detection system for detecting malicious behaviours in big data transaction. In: Information Security and Cyber Forensics (InfoSec) (2015)


3. Goevert, K., Cloutier, R., Roth, M., Lindemann, U.: Concept of system architecture database

analysis. In: 2016 IEEE International Conference on Industrial Engineering and Engineering Management (IEEM) (2016)


4. Johnston, T., Weis, R.: Managing Time in Relational Databases. Morgan Kaufmann,

Burlington (2010)

  5. Kuhn, D., et al.: Expert Indexing in Oracle Database 11g. Apress (2012)

  6. Kvet, M.: Temporal data approach performance. In: APSAC 2015, pp. 75–83 (2015)

7. Kvet, M., Matiaško, K.: Temporal transaction integrity constraints management. Clust.

  Comput. 20, 673–688 (2017)


8. Lee, S., Zeng, X.: A modular method for estimating null values in relational database

systems. In: 2008 Eighth International Conference on Intelligent Systems Design and Applications (2008)


9. Stoja, S., et al.: Architecture of real-time database in cloud environment for distributed

systems. In: 2014 2nd International Conference on Artificial Intelligence, Modelling and Simulation (2014)


10. Qi, C.: Research on null-value estimation algorithm based on predicted value. In: 2014 IEEE

5th International Conference on Software Engineering and Service Science (2014)


New Method for Selecting Exemplars

Application to Roadway Experimentation


Abstract. Nowadays, data are generated and collected in many

domains from various sources. In most of the cases, they are handled

as common data where some simple calculations are used to analyse

them as measuring the average, the maximum, the deviation, etc. For

instance, the average number of children in European families is 1.8 chil-

dren. This kind of assessment is far away from reality: the number of

children should be an integer number. For this reason, exemplars have

a finer meaning since its aim, in this case, is to look of an exemplar of

a common family in Europe which has 2 children (the most represen-

tative family). The aim of this paper is to propose a methodology able

to extract representative exemplars from a dataset. This methodology

has been experimented with dataset extracted from experimentations of

connected vehicle traces. This data analysis has shown some interesting

features: the vehicle connectivity guarantees that messages are not lost.

  the prototypes that are artificial data such as the statistics. Thus, the selection of a few exemplars that represent the whole dataset is one of the first step when exploring a dataset. For instance, the selection of exemplars is central

  . The exemplars are real data extracted from a large dataset unlike

  A set of exemplars is a classical way for storing and representing the cognitive structures

  When we want to study drivers behavior, we have huge amount of data, each corresponds to a behavior. To analyze them, we need to explore these data and we extract exemplars. In order to do this, we propose a new methodology based on an estimation of the local density in a neighborhood of each data. By doing this, we extract representative exemplars that will reflect the different behaviors. We can choose the number of exemplars we want to reduce the processing cost and time. If the classes are known, we try to provide at least one exemplar in each class.

  Data mining Intelligent Transport Systems



  Keywords: Exemplars ·

  1 1 CReSTIC, Universit´e de Reims Champagne-Ardenne, Reims, France


2 LS2N, Universit´e de Nantes, Nantes, France

  Emilien Bourdy

  , and Hac`ene Fouchal


  , Michel Herbin





  , Kandaraj Piamrat


1 Introduction


E. Bourdy et al.

  to several clustering methods

  . The selection of exemplars is a case-oriented

  process which is also called sampling

   ]. The goal is to extract a small subset of representative data from the dataset.

  The use of sampling techniques is necessary when the dataset is too large. But sampling techniques are also used when the treatment of each individual data needs for lot of money, time, effort, etc. Moreover the selection of exemplars is essential in all the fields where tests or treatment are impossible to imple- ment on the whole population (i.e. the whole dataset). In such trials it may be risks associated with individual treatment. So ethical reasons involve testing the treatment on only a small sample of the population.

  Data are described generally with a large number of variables. Because of the sparsity of high dimensional data space, the selection of exemplars becomes highly difficult when data lies within such a space. The phenomenon is known as the curse of dimensionality

   ]. The method we propose uses the approach of

  parallel coordinates

   ] to escape the curse of dimensionality when extracting exemplars.

  The area of VANET is a very challenging domain nowadays. It attracts many research teams mainly to prepare the future vehicles, which will probably be connected and autonomous. Connected vehicles exchange a lot of messages and the need of analysis on these large amount becomes very urgent.


   ], the authors present a formal model of data dissemination within

  VANETs and study how VANET characteristics, specifically the bidirectional mobility on well defined paths, have an impact on performance of data dissem- ination. They investigate the data push model in the context of TrafficView, which have been implemented to disseminate information about the vehicles on the road.


   ], the authors handle two aspects The derivation of real mobility pat-

  terns to be used in a VANET simulator and the simulation of VANET data dissemination achieved with different broadcast protocols in real traffic setting.

  Most of data analysis are done on one hop sent messages but it could be interesting to analyse data over routing issues as the in


2 Sampling Method

  Let Ω be a dataset with n data defined by Ω = {X , X , X , ...X n }.




  3 The goal of sampling is to select a subset of Ω, which is called the subset Σ

  where Σ = {Y , Y , Y , ...Y p } with Y j ∈ Ω.



  3 New Method for Selecting Exemplars Application


  When sampling, p is much smaller than n (p ≪ n) and Y j (with 1 ≤ j ≤ p) is a representative or exemplar of Ω. This paper describes a new method to select these exemplars.

  Our method is based on an estimation of the local density in a neighborhood of each data. The first exemplar we select is the one with the highest local density. Then the nearest neighbors of this exemplar are removed from Ω. We obtain the following exemplars while iterating the process until the dataset is empty.

  2.1 Local Density In this subsection, we explain how we estimate the local density of each data and how we define the nearest neighbors of an exemplar. Finally, we study the number of exemplars we can propose using this sampling method.

  In this paper we only consider multidimensional quantitative data. Thus, X i with 1 ≤ i ≤ n is a vector defined by: X i = (v (i), v (i), ...v p (i))




  where v

  1 , v 2 ,... v p are the p variables that are the features of data. In this context each data lies a p-dimensional data space.

  Classically, the density is defined using a unit hypervolume. For instance, the hypersphere of radius α can define the unit hypervolume. In the data space, the local density at X is then equal to the number of data of Ω lying inside the unit hypersphere centered in X. Unfortunately the definition of density comes up against the curse of dimensionality

  . When the dimension of the data space

  increases, the volume of the available data becomes sparse and the classical definition of density has no meaning. For circumventing this drawback, we define the density for each variable using the approach of parallel coordinates

  . Therefore, we have p densities, each defined in a one-dimensional space.

  The sum of these densities gives us a density-based index that we use in the whole data space.

  Let us define the density computed in the one-dimensional space of the vari- able v j (with 1 ≤ j ≤ p). The dataset Ω is projected in this space and we obtain n values with: Ω j = {v j (1), v j (2), v j (3), ...v j (n)}.

  These values are in the range [min j , max j ] where min j = min (v j (i)) and


  max j = max (v j (i)). Let us define the unit interval we use to compute the


  density at each value x. Let k be an integer between 1 and n. If we expected a local density equal to k, then the length α j we propose for the unit interval is max −min j j equal to α j = ∗ k. Thus the local density at x is equal to the number n of elements of Ω j that are in the unit interval [x − α j /2, x + α j /2]. The local density at X i for the variable v j is then defined by: density j (X i ) = #{ [v j (i) − α j /2, v j (i) + α j /2] ∩ Ω j }.


E. Bourdy et al.

  Finally, the local density at X i for all the variables is defined by: density(X i ) = density j (X i ).



  We select the data which has the highest local density. This data is the first exemplar of Ω: Y

1 = arg max density(X i ).

X i ∈Ω

  2.2 Nearest Neighbors The previous procedure enables us to select only one exemplar. We obtain the following exemplars by reducing the dataset and iterating this procedure. The dataset is reduced by removing Y 1 and its nearest neighbors.

  Let us describe our definition of the nearest neighbors of a data X in a dataset Ω. The neighbors of X i for the variable v j are the data of Ω that are in the unit interval centered in X . This neighborhood N is defined by: i j N j (X i ) = {X k ∈ Ω with v j (k) ∈ [v j (i) − α j /2, v j (i) + α j /2]}.

  The nearest neighbors of X i for all the variables should be in the neighborhoods for each variable. Thus the nearest neighbors of X i are in the neighborhood N defined by: N (X i ) = N j (X i ).



  To select the second exemplar Y we exclude the first one Y and its nearest



  neighbors N (Y

  1 ). We apply the procedure defined in the previous section within

  a reduced dataset Ω \ N (Y

  1 ). Then Y 2 the data with the highest local density within the reduced dataset.

  We iterate the procedure until the reduced dataset is empty. The exemplars we obtain gives us the samples of Ω.

  2.3 Number of Exemplars We set our method of sampling using the parameter k where k is an expected local density at each data. The value of k lies between 1 and n when the dataset has n data. In this section, we explain how the value of k can change the number of exemplars selected through our sampling method.

  Let us consider a toy example with 200 simulated data (n = 200) with 5 variables (p = 5). Figure

   displays the profiles of these data with 200 dashed

  broken lines. The exemplars are selected using the parameter value k = 100. We obtain 7 exemplars (bold broken lines in Fig.


  The number of selected exemplars decreases when the parameter value k increases. Figure

   shows that the number of selected exemplars decreases from

  200 to 1 when the density parameter k increases. This property of our method

  New Method for Selecting Exemplars Application

  79 Fig. 1.

  Profiles of 200 simulated data with 5 variables (dashed lines) election of 7 exemplars with a parameter value k = 100 (bold lines) Fig. 2.

  Number of selected exemplars decreases from 200 to 1 when the density param- eter k increases from 1 to 200

  enables us to adapt a strategy to select the number of samples that we extract from the dataset. If we want a specific number of samples selected from the initial dataset, then we can adjust the parameter k to obtain the expected number of exemplars.

3 Assessment of Sampling

  The exploratory analysis of a dataset is complex for many reasons. The dataset is often divided into classes but the distribution of these classes is unknown.


E. Bourdy et al.

  Moreover, the number of these classes is also unknown. To better understand data, the use of an complementary exploratory trial on a smaller dataset is often necessary. The selection of a reduced number of samples should then represent all the classes of the dataset. For this reason, we will evaluate our sampling method under controlled conditions when the distribution of the classes is known. But of course, the method remains designed for applications in exploratory analysis when the classes are unknown. This method is particularly useful when classes have large overlapping and when the classes have very different numbers of data. In such cases, the classical methods of clustering very often fail.

  Let us consider a dataset with known distribution of classes for assessing our sampling method. We verify that the distribution of the selected exemplars between classes remains comparable with the distribution of data within the initial dataset. Table

   gives the results we obtain with some simulations.


Table 1. Distribution between classes within a dataset (n = 200) and within the

selected exemplars Number of classes Distribution in Number of Exemplars dataset (n = 200) selected distribution between exemplars classes 4 (42, 51, 65, 42) 25 (4, 8, 8, 5) 4 (42, 51, 65, 42) 18 (5, 6, 4, 3) 4 (42, 51, 65, 42) 13 (3, 3, 5, 2) 4 (42, 51, 65, 42)

  9 (2, 3, 3, 1) 4 (42, 51, 65, 42) 7 (2, 2, 2, 1) 4 (7, 103, 62, 28)

  10 (1, 3, 4, 2) 5 (40, 47, 55, 9, 49) 10 (2, 3, 2, 2, 1) 6 (6, 12, 76, 80, 24, 2)

  9 (2, 1, 2, 1, 2, 1) 7 (23, 16, 51, 46, 1, 8 (1, 1, 1, 2, 0, 1, 2) 36, 27)

  8 (37, 9, 3, 19, 48, 12, 9 (3, 0, 0, 1, 1, 1, 3, 1) 45, 27)

  In the first five rows of the table, we use the dataset displayed in Fig.


  dataset is simulated using four classes with a large overlapping. The 200 data are randomly distributed between these classes. (42, 51, 65, 42) is the distribution between the four classes. The number of selected exemplars decreases when the parameter k increases. We obtain 25, 18, 13, 9 and 7 exemplars using respectively 50, 60, 70, 80 and 100 as values of k. In these five simulations, the four classes are effectively represented by the exemplars. However, when k increases, the number of selected exemplars becomes too small for representing each class.

  In the last five rows of Table

  we simulate five datasets with respectively

  4, 5, 6, 7 and 8 classes. The number of data in each class is randomly selected

  New Method for Selecting Exemplars Application


  and it could be very different from one class to another one. The datasets have 200 data and the parameter k is equal to 80 when selecting exemplars. When the number of classes increases, the number of exemplars becomes too small for representing each class (see the two last rows of the table). However, these classes are represented if the number of selected exemplars increases (i.e. if we decrease the value of the parameter k).

  Let us study the sampling with real datasets. We consider some datasets of UCI repository (see in

   displays the selection of exemplars using our

  blind method (i.e. when the classes are unknown) on the classical dataset called “Iris”, “Wine”, “Glass”, “Haberman” and “Ecoli”.


Table 2. Distributions between classes with a real dataset and with selected exemplars

(n = number of data, p = number of variables)

Name of dataset n p Distribution in Number of Distribution of

dataset exemplars exemplars Iris 150 4 (50, 50, 50) 8 (3, 3, 2) Wine 178 13 (59, 71, 48)

  9 (2, 4, 3) Glass 214 9 (70, 76, 17, 13, 19 (1, 6, 1, 5, 1, 9, 29) 5)

  Haberman 306 3 (225, 81) 10 (7, 3) Ecoli 336 7 (143, 77, 2, 2, 23 (3,9,0,0,3,3,2,3) 35, 20, 5, 52)

  These datasets have respectively 3, 3, 5, 2 and 8 classes. Our sampling method gives generally an exemplar in each classes. Obviously the method fails if the number of classes is high relative to the number of selected exemplars. Moreover, the method often fails if the number of elements within one class is very low. For instance, in the last line of Table

  two classes have only 2 elements and

  these classes are not represented by the exemplars. But these classes can be represented by an exemplars if we increase the number of exemplars we select.

4 Roadway Experimentation

  In the Scoop@f

   ] (Cooperative System at France) project, Intelligent Transport

  System (ITS) is experimented in the real life. To do that, connected vehicles drive on roadway and communicate with the infrastructure or other vehicles via a spe- cific WiFi called ITS-G5. Messages used in Scoop@f are CAM

   ] (Cooperative

  Awareness Message) and DENM

   ] (Decentralized Environmental Notification

  Message). CAM is an application beacon with information about ITS station position, speed if it’s a mobile station, etc. DENM is used to warn about events. In this experimentation, the vehicle drives on a roadway and send DENMs auto- matically. The event of this experimentation is a slippery road. When vehicle


E. Bourdy et al.

  send DENM, it logs the 30 previous seconds and 30 next seconds. We used theses logs with our methodology.

  Theses logs contain 3 201 data of 17 variables (7 for the acceleration control, the steering wheel angle, the strength braking and 8 for the exterior lights). The acceleration control is defined by the brake, gas and emergency brake pedals, collision warning, ACC (Adaptive Cruise Control), cruise control and speed lim- iter utilization. And the exterior lights are defined by the low and high beam, left and right turn signal (warning is the combination of both), daytime, reverse, fog and parking light.

  Here we want to describe characteristics from the experimentation and trying to modeling vehicle behavior on this type of route. By using the 3 201 data we obtain the Fig.

  we can see that there no big differences. It’s explain by the fact

  that roadway has less changes than urban road. We then used our methodology with the k ∈ {40, 80, 100, 140, 180, 200} that give us the Table

  With k = 40,

  142 samples are extracted, 102 with k = 80, 92 with k = 100, 58 with k = 140, 46 with k = 180 and 48 with k = 200. The reduction of number of samples in comparison with the number of entries is explained by the fact that there is a lot of data that are equals. With our methodology, a big pre-processing is made, dividing by {22, 31, 34, 55, 69, 66} the number of data to process.

  Fig. 3.

  Profiles of 3 201 data with 17 variables from roadway experimentation. Table 3.

  Selection of exemplars with different values of the parameter k from roadway experimentation.

k Number of exemplars Division

40 142

  22 80 102 31 100


  34 140


  55 180


  69 200


  66 New Method for Selecting Exemplars Application


5 Conclusion and Future Work

  project to perform exemplars of the situation described by the experimentation.


5. Bache, K., Lichman, M.: UCI Machine learning repository, School of Information

and Computer Sciences, University of California, Irvine. (2013)


10. CAM: ETSI EN 302 637–2; Intelligent Transport Systems (ITS); Vehicular Com-

munications; Basic Set of Applications; Part 2: Specification of Cooperative Aware- ness Basic Service. European Standard. ETSI, November 2014

  9. Scoop@f.

  VANETs. In: 2013 IEEE International Conference on Communications (ICC), pp. 1424–1428 (2013)


8. Ayaida, M., Barhoumi, M., Fouchal, H. Ghamri-Doudane, Y., Afilal, L.: PHRHLS:

a movement-prediction-based joint routing and hierarchical location service for


7. Castellano A., Cuomo F.: Analysis of urban traffic data sets for VANETs simula-

tions. CoRR abs/1304.4350 (2013)


6. Nadem T, Shankar P, Iftode, L.: A comparative study of data dissemination models

for VANETs. In: 3rd Annual International Conference on Mobile and Ubiquitous Systems (MOBIQUITOUS), July 2006

  (eds.) SSDBM 2010. LNCS, vol. 6187, pp. 482–500. Springer, Heidelberg (2010).

  In the near future, we will use the methodology with other experimenta- tion and developing tools to create classes and representing each experimenta- tion. This work is a contribution for designing tools to analyze data of roadway experimentations in the project Scoop@f.

  In this paper, we have presented a new methodology to select exemplars from a dataset containing multidimensional quantitative data. Our method is based on an estimation of the local density in a neighborhood of each data. With this methodology, it’s also possible to select exemplars from classes, and then reduce the number of data in theses classes. This methodology was first used with ran- dom data and then with known real data, and with a roadway experimentation from the Scoop@f


3. Cochran, W.G.: Sampling Technique. Wiley Eastern Limited, New Delhi (1985)


2. Frey, B.J., Dueck, D.: Clustering by passing messages between data points. Science

315 , 972–976 (2007)


1. Frixione, M., Lieto, A.: Prototypes vs exemplars in concept representation. In:

International Conference on Knowledge Engineering and Ontology Development, KEOD (2012)



Acknowledgement. This work was made possible by EC Grant No. INEA/CEF/

TRAN/A2014/1042281 from the INEA Agency for the SCOOP project. The statements

made herein are solely the responsibility of the authors.


4. Houle, M.E., Kriegel, H.-P., Kr¨ oger, P., Schubert, E., Zimek, A.: Can shared-

neighbor distances defeat the curse of dimensionality? In: Gertz, M., Lud¨ascher, B.


E. Bourdy et al.


11. DENM: ETSI EN 302 637–3; Intelligent Transport Systems (ITS); Vehicular Com-

munications; Basic Set of Application; Part 3: Specifications of Decentralized Envi- ronmental Notification Basic Service. European Standard. ETSI, November 2014


12. Heinrich, J., Weiskopf, D.: State of the art of parallel coordinates. STAR? State

of The Art Report, Visualization Research Center, University of Stuttgart, Euro- graphics (2013)


Temporal Flower Index Eliminating Impact

of High Water Mark


  Michal Kvet and Karol Matiasko


Faculty of Management Science and Informatics, University of Žilina,

Univerzitná 8215/1, 010 26 Žilina, Slovak Republic

Abstract. A huge amount of data is produced daily, which should be managed,

handled and stored in the database. Intelligent system characteristics are mostly

based on storing data over the time with regards to validity. The core of the

system is therefore framed by the temporal database, which offers the possibility

for data analysis, decision making or creating future prognoses. None of the data

should be stored indefinitely, effective data management is an inevitable

part. This paper references historical background and temporal evolution with

emphasis on various granularity modeling and managing describes index

structure types used in database approaches covering access paths taken out by

optimizer techniques. The main contribution of this paper is Flower Index

Approach, which aim is to eliminate the impact of database High Water Mark if

Full Table Scan access method is used. Thanks to that, we can optimize costs,

reduce processing time and increase performance, which is also supported by

experiment comparisons. Keywords:

Column level architecture Flower Index Approach

Temporality Volatility Full Table Scan Index forcing Data fragmentation

1 Introduction – Temporal Environment Characteristics Development of any information system is influenced by the data storage management.

  Increasing number of parameters and data complexity significantly forces the pro- grammer to take care of effectivity. Looking through the history, it can be concluded, that data themselves should be shifted into specific storage – database. Current database systems are based on relational database covered by transactions highlighting only current valid states

   ]. Relational systems are still available, powerful, but limited in terms of time. Temporal database approach is based on managing tuples over the time.

  In the past, it used backup and log files to provide a historical image. Later object level temporal architecture was introduced based on extending primary key of the system with the time element. Each update therefore composed new state. If some data portion was not changed, the particular value was copied to new state degrading performance. The column-oriented approach is based on attribute granularity, which is suitable for sensorial data with various granularity and frequency of changes. In 2017, the hybrid solution was proposed by us. In such case, granularity is based on attribute groups, which are detected and managed automatically. Historical background and evolution

86 M. Kvet and K. Matiasko

  One of the most significant parameters influencing usability, effectivity and overall value of the solution is just performance of the query, which can be bounded by several environment properties and system structures. Temporal systems are mostly high- lighting data retrieval process and update operations. Their architectures consist of several levels, where performance optimization can be placed [

  In this paper, we

  concern on index approaches, access paths and issues specific to them based on data fragmentation management. The index managed data are immune to physical frag- mentation in data files.

  Index structures can be defined as optional structured segments associated with a table, which aim is to speed data retrieval, to reduce read operations from the disk. Commonly, tables are organized in heap, thus if no another layer consisting of indexes is defined, Full Table Scans is performed to find appropriate values by scanning all associated blocks of the table. Presence or absence of the index does not influence code of the SQL, but the performance can be influenced rapidly. Accessing data using index is optimal for 15 –25% of all table rows.

  Information about defined indexes can be found in multiple data dictionary views like {user | all | dba}_indexes, {user | all | dba}_constraints (for dealing with con- straints involving index definition). List of indexed columns and the order of attributes in the index can be found in {user | all | dba}_dict_columns. Size of the segments (index, table) is accessible by querying {user | all | dba}_segments data dictionary view. Whereas indexes are extensive for a complex temporal table, they are not usually backed up, whereas no data can be lost (all are still stored in data files ensured by transaction management and backups, restore and recovery politics). Code definition of the index can be obtained by GET_DDL function of the SYS owned package DBMS_METADATA with two parameters

  • – object type and the name of the object itself [ ].

  Default database index type is B-tree, respectively B+tree, which is mostly used because it maintains the efficiency despite frequent changes of records (Insert, Delete, Update ). B+tree index consists of a balanced tree in which each path from the root to the leaf has the same length.

  Limitation of this approach is a small number of records (low cardinality). In that case, using index does not produce the desired effect in terms of performance (accel- eration). Another disadvantage is the lack of support SQL queries with functions implicitly.

  There are also other index types, which are mostly delimited by specific approaches and application domains

  • – bitmap, hash, reverse key, function-based or domains. A special category is a cluster [

  In the relational database management system,

  access path refers to the method of data retrieval chosen by optimizer based on requested SQL statement. Generally, optimizer makes an independent decision based on heuristical approaches, which input values are mostly statistics gathered by back- ground process MMON

  • – Manageability Monitor. This step is very important because it influences the principles and methods of data retrieval. There are two basic types of access paths:

  • Full Table Scans, • Index Access Paths.

  Temporal Flower Index Eliminating Impact of High Water Mark


  Index Access Path principles are based on using the index to accelerate row location. Generally, it uses two branches

  • – index column values (or function result based on index columns) and ROWIDs, if necessary. Whereas default defined index is B+tree, processed values are sorted automatically using the linear linked list in the leaf layer.

  There are several methods of the Index Access Paths (Index Unique Scan, Index Range Scan, Index Skip Scan introduced in Oracle 9i, Index Full Scan, Index Fast Full Scan

  , etc.), which are described in detail in


  In the next chapters, we will deal with data management over the time with regards to fragmentation causing performance degradation, when the suitable or usable index is not defined.

2 Existing Solutions

  All information systems based on temporal databases are covered by indexes to opti- mize data retrieval. Vice versa, it slowdowns the performance of other Data Manip- ulation Language (Insert, Update, Delete) statements due to the need for index actualization and balancing. It is not possible to develop all possible indexes based on queries, which can moreover evolve over the time, as well. Full Table Scan method itself is mostly considered as the severe and demanding operation to get data from the database. Data are not accessed by the index and locators (ROWIDs) but are sequentially read using all rows.

  Several kinds of literature [

   ] deal with reading all rows, which is essentially

  correct but also greatly simplified in the terms of performance. Data rows, nor blocks, are linked together, so it is very complicated, even impossible, to locate next data tuple. Thus, all data blocks belonging to the particular table must be loaded, accessed and sequentially processed. Even if the table is small consisting only a few blocks, data rows can be diffused to several blocks, some blocks can be even empty, which cannot be, however, evaluated by background processes without accessing. Index hint does not solve the problem, the optimizer does not need to use it. Moreover, index forcing provides a worse solution in comparison with direct optimizer solution. Optimizer decision is based on statistics and available indexes. First of all, index list is loaded searching suitable on. If found, it is used to access data, otherwise, Full Table Scan method is used.

  During the data management, data are loaded, updated and deleted resulting in data fragmentation. Actual approaches, principles, and concepts of dealing with data storing efficiency and fragmentations are based on two separate streams for data management – online and offline defragmentation with their improvements.

  Offline data defragmentation approach is considered as the most effective solution because after executing such operation, no data fragmentation is present. The simplest way is to delete all data accompanied by new data loading. The main disadvantage of such approaches is just locking operation. Thus, the processed table must be put into offline mode and cannot be accessed. However, it cannot be done directly using table granularity, but the whole tablespace must be indicated as offline. It results in two

  88 M. Kvet and K. Matiasko

  other segments are moved into free and not processed tablespace. Anyway, data of the particular table will not be accessible during processing, which is a significant limi- tation of the commercial solution. Some improvements have been done based on data tuple size and merging operations


  The online solution is the second group. In that case, the whole table is not locked, but the key lock is applied only for the block itself or to the extent. Management methods store sorted the list of the tuples size, which can, however, exceed block size. In that case, a reference to the next block recording the tuple must be stored. If there is free space in the block, First Fit Method or Best Fit Method is used. Although the solution provides suitable results, it should be remembered, that it requires a large number of system resources as well and memory space for storing the list [


  Therefore, it is often performed during the off-peak time period, which sometimes degrades into offline solution. However, the difference is based on availability.

  It would seem that described solutions are effective and performance efficient, which is not, however always true. As it has been already mentioned, temporal systems obviously manage complex data during the whole life cycle, which generates signifi- cant data amount and requires great demands on resources. We propose own solution, which is not based on defragmentation but uses special index type – Flower Index Approach with regards to Full Table Scan method.

  Three most significant factors are covered by our proposed solution:

  • index management,
  • data volatility, • data structure degradation.

  3 Own Solution

  The role of the temporal system is a comprehensive data management over the time, to store object states existing in the past, present, as well and planned condition changes. Processed data amount rises significantly, however, the issue of storing too old data does not reflect benefits, obviously. The temporal approach is used for actual decision making based on definite time frame and conditions, which can vary on each attribute (can be even infinite for some attribute sets). Therefore, it is necessary to distinguish useful data from expired, which can be moved to another data repository mostly modeled by archives or data warehouses. The problem of incorrect data management is not covered in the temporal system, therefore we propose own methods for dealing with non-actual states with regards to their identification and performing the appro- priate actions.

  Based on application analysis, the decision on a time frame or number of never states can be executed. Volatility can be expressed by losing weight and relevance of the attribute values over the time, if applicable. It can happen, that long-term invalid data are not necessary to be stored in the production database and reachable online immediately (notice, that all data moved from the active database can be found in another repository).

  Temporal Flower Index Eliminating Impact of High Water Mark


  A special type of the volatile attributes is produced by security management of sensitive data

  • – strictly defined period of time for data storing can be defined and imposed, so after the interval has been flown, appropriate data must be removed completely, covering also all data repositories. These constraints are mostly applicable to security systems, medical information systems or bank systems determined by actual laws. For temporal approaches, we have defined three methods for removing, respectively transforming volatile data: • Purge – delete data expired before defined time point.

  • Delete_count – data removal is based on the defined number of newer states detached to defined attribute or attribute set.
  • Delete_time – reflects the time interval for storing defined values. It can be handled by begin or end point of the validity.

  In our approach, each of proposed methods can be defined in multiple data granularity:

  • table,
  • object, • attribute.

  Performing volatility operations can significantly influence access methods, which can have significant impacts on performance. Data are moved away resulting in freeing disc space and creating data fragmentation, which is not managed automatically. Thus, during the processing phase, physical disc space is not released. Historical records removal and transfer from the structure involves the creation of free space within the block and its fragmentation if it is not possible to full it completely. Most data in temporal databases are changed really frequently. After an even short time, there are significant performance impacts and slowdowns, if Full Table Scan method is used. However, as it has been mentioned, it is not possible to get rid of it completely. The full Table Scan method right border is based on High Water Mark (HWM), which is top limit for the number of blocks that have been used for the table [

   ]. HWM value is

  stored in table segment. All blocks allocated for the particular table must be accessed to control, whether they contain row or not. Temporal databases are characterized by high workload – a significant stream of changes over the time with regards to removing expired data (based on volatility). Mostly removing, respectively transferring historical data to another data location is performance significant. HWM value is not moved backward in case of performing Update, Delete or Purging statements. Thus, as a consequence Full Table Scan is delimited by HWM value, which is in a standard condition still increasing. To solve the problem, we propose our own solution based on Flower Index Approach (FIA), which handles blocking factors, evaluates a number of empty blocks allocated for table segment. Although these values are not collected during statistics gathering, there is still a possibility to get relatively small frame covering exact value. It evaluates a number of blocks under HWM, which can be obtained from user_tables data dictionary view (DDV) accessing blocks attribute. A number of rows in the table is part of the statistics – attribute num_rows.

  The last parameter value for decision making is the average length of the row in the

90 M. Kvet and K. Matiasko

  DB_BLOCK size initialization parameter specifies the size of the database block. Typical values are 4096 and 8192 bytes (default value is 8192 bytes).

  Blocking factor is therefore expressed as the proportion between the used size of the physical storage and allocated space. Even if the blocking factor has really low value (e.g. 0.1), when Full Table Scan is used, all blocks are accessed and controlled for row values. However, it means that 90% of processing is a waste time.

  Our proposed solution is based on forcing the system to use index regardless the input estimate time and performance. It highlights blocking factor as a more important performance parameter.

  Database instance guarantees, that each block is stored no more than once in the buffer cache. Thus, management tables (in principles, it can be stored in data dictionary views, but a possible solution is also user-defined table managed by SYS user) should be extended by one attribute for each table consisting of the number of blocks in the table for a particular segment. Thanks to that, we can evaluate the impact and costs of loading and access data using the index. Figure


  shows the UML flowchart diagram of Flower Index Approach avoiding Full Table Scans. First of all, system checks, whether there is any index for the table. If so, and some of the indexes are already in the memory (buffer cache), locks are applied, which ensure, that blocks of the mentioned index will not be removed from the buffer cache, if there is need to get free space (it is marked similar to dirty buffer block). Thus, if all index blocks are memory access, data retrieval using such index can be performed. If there is only part of the index in the memory, missing blocks are loaded from data files. Whereas memory index blocks are locked, they cannot be overwritten, thus, as a consequence, after such operation, entire index is loaded in the memory. We assume that size of the buffer cache memory is sufficient to store entire index if not, swapping technologies would be used.

  A special situation occurs if no index is in the memory. In that case, all blocks of the index must be loaded. This operation is usually still faster than accessing all blocks of the table. Be aware of temporal approaches managing significant data amount over the time.

  The last step is based on accessing rows using index ROWIDs. It can be said, that index is not used for searching based on index columns (Where clause), but is mostly used for locating rows in the table segment. After the operation, locks are released.

  Naturally, if there is no index in the system (which is against principles of the relational theory due to primary key definition), there is no another way, Full Table Scan must be used.

  Flower Index Approach is delimited by the initialization parameter al- low_FIA_index , which can have three possible values: • Manual – using FIA principle must be coded explicitly for the query.

  • Force – each time, Full Table Scan would be used, it is replaced by FIA, if applicable (any table index is defined).
  • None – never uses FIA.

  Parameter allow_FIA_index can be set for session and system granularity:

  Fig. 1.

  Flowchart - FIA evaluation Temporal Flower Index Eliminating Impact of High Water Mark


  92 M. Kvet and K. Matiasko

  If the user wants to use FIA for the manual value of the allow_FIA_index parameter, such requirement must be announced by hint mark. The hint can contain also the name of the index to be used. In that case, the index choosing evaluation is not executed, user decision is used. Be careful, when selecting it, automatical proposed method usually makes it better, because it is based on actual buffer cache and loading information. The second principle defines the only hint without index definition itself, thus optimizer automatically selects the best one based on statistics and heuristical methods:

  The last parameter influencing FIA approach is the limit value of the blocking factor. As we will see in the experiment section, proposed solution effectivity is based on blocking factor value. Parameter FIA_limit stores the limit for use proposed tech- nology. If the value is higher, an original approach using Full Table Scan is used. The default value is 70% and can be changed for session and system independently:

  4 Performance

  High workload, online management, and processing require data accessible in unlim- ited measure. Thus, Full Table Scans perform still worse and worse, if multiple data update, insert and delete operations are performed. On the other hand, index approa- ches are not influenced by that fact. Although it can be significant demands on disk space, it is not mainstream. Blocks can be grouped and space shrinking, which means moving tablespace offline and rebuild table to remove fragments and unused space is really special occasion executed almost never. The main reason focuses on the com- mercial sphere, where data must be still available. Thus, offline fragmentation methods are not compared with our solution.

  The challenge is thus to eliminate the need for sequential block scanning or to eliminate fragmentations. One of the possible solutions, but mostly primary-process, is to define the new index. However, it is necessary to highlight, that it is not possible to create any combination of columns in the index. Moreover, index management is not free. Flower Index Approach is located in the first phase of the data access retrieval method delimited by the optimizer. This section will evaluate benefits and limitations of the proposed solution.

  Experiment results were provided using Oracle Database 11g Enterprise Edition Release – 64 bit Production; PL/SQL Release – Production. Parameters of used computer are:

  • Processor: Intel Xeon E5620; 2,4 GHz (8 cores),
  • Operation memory: 16 GB,

  Temporal Flower Index Eliminating Impact of High Water Mark


  Experiment characteristics are based on real environment consisting of 1000 sen- sors producing data ten times for one second. If the difference between consecutive values is lower than 1%, such values are not stored in the database and original value is considered as unchanged. Thus, based on our environment, the average amount of new values is approximately 1000 per second. Amount of data after one hour is 3 600 000. The system is stored in the column level temporal system (comparison of column level and object level temporal system can be found in

   ]). Volatility each object state is

  1 min (which is started after the real change

  • – the difference is higher than 1%). Figure

   shows the results of the experiment processing. There are 3,6 millions of

  rows in the temporal table after the change. When dealing with all data (regardless the volatility), retrieval lasts 52 s and requires 3447 costs of CPU. However, when volatility is applied (each value is usable during 1 min after update), the performance of the Full Table Scan method degrades significantly. In principles, after one processing hour, there are approximately 60 000 of valid and volatile rows (approximately 1,7%). If there is no suitable index

  • – Full Table Scan requires almost the same amount of system resources – costs 3212 and processing time 51 s, which reflects the improvement 6,82% for costs and 1,92% for processing time. The only difference, when comparing processing steps, is moving data to output result set based on condition, which is, however, evaluated during the last step – all data are loaded and evaluated. Our defined Flower Index Approach forces system to use the index. A number of rows to be processed in the result set is the same. Significant improvement is reached in costs – 930 (3% CPU) and processing time


  • – s, which reflects the improvement (in comparison with Table Access Full with expired data – reference 100%) - 71,05% for costs and 64,71% for processing time.


Fig. 2. Experiment results (expired data in the table segment)

  When dealing with online rebuilding methods, two types have been compared with our FIA approach. The first system is based on defragmentation using blocks (block_defragmentation), the second one uses extent_defragmentation. For both

94 M. Kvet and K. Matiasko

  approaches, we compare performance, where rebuilding operation is executed before, respectively during the execution of the Select statement. In all cases, we get better results.

  Block_defragmentation method requires more processing time and system resour- ces because it tries to group each block, whereas extent_defragmentation is charac- terized by extent management, which consists of multiple blocks. Thus, the data amount in case of extent_defragmentation is lower.


   shows the performance results of the system dealing with expired data

  directly in the table segment. Shortcut TAF reflects Table Access Full (synonym to Full ) method, FIA expresses Flower Index Approach.

  Processing time and costs in the table are separately evaluated if defragmentation and data retrieval itself is managed separately. Number of data to be processed provides the best value for FIA approach (6M2 bytes), then, it is followed by the block defragmentation method (7M1 bytes, 7M5 bytes), whereas it consists of less number of free space in blocks in comparison with extent managed defragmentation. Costs of TAF after block defragmentation is 1140 (slowdown of 22, 58% - reference model FIA (100%)). However, it must be highlighted, that another operation – block defragmentation

  • – had to be executed sooner with 1665 costs. Extent management provides better performance of defragmentation itself, however, consecutive data retrieval is worse (costs: 1198 (slowdown 28,82%), time: 22 s (slowdown 22,22%) reference model: FIA (100%)).

  Slight difference between TAF and FIA can be perceived in the number of bytes (8,77%). The reason is based on index management and loading. Minimal part can be also influenced by row migrating after an update operation.

  Compared to the most appropriate index for this query (based on Where condition), the performance of FIA for Select statement is worse using approximately 2% (upper part of the Fig.


  Fig. 3.

  Experiment results (expired data are moved automatically)

  Temporal Flower Index Eliminating Impact of High Water Mark


  Another system moves expired data to the archive repository. It can be done either automatically (it is not necessary to deal with it explicitly, however, it can influence performance during peaks) or manually (it is executed in the service window during lower database traffic). In our experiments, we use automatic management (expired data are moved immediately), although manual solution has lower impacts on per- formance. The reason is based on setting properties and definition of the workload strength. Performance results are shown in lower part of the Fig.


  Full Table Scans do not produce performance improvements if data are moved to another repository, whereas blocks are not deallocated. Moreover, newly inserted values can even be not located to empty part of the existing blocks, but new blocks would be allocated (existing empty blocks are used only for an extension after updates). FIA performance (when data are moved) is better (approximately 5%), because of the index

  • – the index is loaded to the memory buffer cache, then, it is used for evaluation resulting in loading, if there is only part of the index in the memory or even no part. Thus, the first step the faster approach of data retrieval. Notice, that FIA approach does not need to extend index definition, thus it does not have any impact on destructive DML operations.

  In comparison with the previous experiment, processing time, costs and number of processed bytes are lowered. It is the result of reducing the size of the index structure. For all experiments, the only primary key of the table (ID_change) has been used. No additional index structures have been developed.

  The previous part of the section deals with the retention policy of the value volatility set to 1 min. If the policy is unchanged, after a longer time interval, the ratio between allocated blocks and real useful data will be still more significant (after one hour, we have approximately 2% of unexpired data, after two hours, only 1%, after the day, there is only 0,07% of unexpired data). Therefore, term blocking factor has been introduced, which expresses the ratio between expired (empty) rows stored in the data files in comparison with valid unexpired data.

  It is necessary to set bounds of expiration for using FIA in comparison with Full Table Scans

  . Whereas it requires index management and loading, if the blocking factor is high, FIA reaches worse results. Figure

   shows the performance characteristics and

  results for multiple blocking factor values with regards on reduction factor impact. The number of data and conditions are the same as the previous experiments.

  96 M. Kvet and K. Matiasko

  In terms of performance, it can be said that FIA approach is more appropriate to a limit of 70% for blocking factor (TAF costs are 3447 and processing time 52 s). Vice versa, with higher values, original Full Table Scan methods are preferable, there is no index loading and evaluation necessity.

  Our proposed FIA approach deals with data fragmentation by locating data using index structure, whereas referenced defragmentation methods remove free space and groups blocks or extents based on defined approach granularity. The best solution provides the combination of the systems, that it is possible to identify a period of a low load of the system (usually at night). Existing defragmentation processes are auto- matically planned to maintenance window. Thus, after such operations, the system is fully optimized with emphasis on free blocks. However, between those intervals, a large number of operations and volume of data are processed, the structure is altered and new data images are created. Again, fragmentation is created, but cannot be managed by defragmentation operations immediately because of multiple performance limitations – required resources and effectivity. In that case, therefore, our FIA approach is used. As we can see, such solution improves performance using 8% rate. Block defragmentation (planned every hour). FIA is always used after half an hour after the end of defragmentation. Expiration settings and parameters are the same as in the first experiment.

  Method FIA (1 -min expiration) required 930 for costs, 3% of CPU and 18 s as processing time. Method block defragmentation + FIA required 856 for costs, 3% of CPU and 17 s as processing time.

  5 Conclusion

  Development, usability, and performance of the information system are significantly influenced by the data approaches. Data evolve over the time and must be stored in the database, whereas they provide input data for analysis, decision making, optimization and monitoring activities. In the first part of the paper, we reference multiple approaches, which cover temporality with regards to granularity and frequency of changes. The column-level approach reduces duplicates, which can cause transaction problems and inconsistency.

  Historical data can lose their relevance and they should be removed from the production database after passing defined conditions. Therefore, we propose methods for dealing with volatility. Performance optimization must be therefore provided in multiple spheres. First of all, the temporal structure must reflect the strength and properties of processed data with emphasis on physical structure. Fast and reliable access is the core part of the intelligent information system as well as provided per- formance. Optimizer automatically evaluates defined index to choose the best method for data access. In this paper, we highlight Full Table Scan method, which is based on accessing all blocks covered by the table definition. Whereas they can be empty or partially empty, performance degrades. Therefore, we propose our own solution created by index definition and data locating by Rowids – Flower Index Approach. Flowchart

  Temporal Flower Index Eliminating Impact of High Water Mark


  Our approach is complexly evaluated in the experiment section, where performance and limitations are defined. The best solution of existing approaches reaches block defragmentation. In that case, execution costs are lowered using approximately 18% for costs and 10% for time using FIA approach. Then, our proposed approach is also evaluated with best suitable index structure. Naturally, our approach must be worse, but the slowdown in performance is minimal compared to the index administration costs −2,4% for costs and 5,5% for a time. Then, reduction factor impact is compared. In the last part, we deal with the combination of existing approaches extended by FIA approach, where 8% rate improvements can be reached.

  In the future, we would like to introduce this solution to cluster and replication environment, to deal with index distributions and definition of the new expert layer for dealing with partially free blocks to locate new rows.

  Acknowledgment. This publication is the result of the project implementation:

Centre of excellence for systems and services of intelligent transport, ITMS 26220120028

supported by the Research & Development Operational Programme funded by the ERDF and

Centre of excellence for systems and services of intelligent transport II., ITMS 26220120050

supported by the Research & Development Operational Programme funded by the ERDF.







1. Aibin, M., Walkowiak, K.: Defragmentation algorithm for joint dynamic and static routing

problems in elastic optical networks with unicast and anycast traffic. In: 2016 International Conference on Computing, Networking and Communications (ICNC) (2016)

2. Arora, S.: A comparative study on temporal database models: a survey (2015)


3. Ahsan, K., Vijay, P.: Temporal Databases: Information Systems. Booktango, Bloomington



4. Claramunt, Ch., Schneider, M., Wong, R.C.-W., Xiong, L., Loh, W.-K., Shahabi, C., Li, K.-

J. (eds.): SSTD 2015. LNCS, vol. 9239. Springer, Cham (2015).


5. Chomicki, J., Wihsen, J.: Consistent query answering for atemporal constraints over

temporal databases (2016)


6. Goevert, K., Cloutier, R., Roth, M., Lindemann, U.: Concept of system architecture database

analysis. In: 2016 IEEE International Conference on Industrial Engineering and Engineering Management (IEEM) (2016)


7. Johnston, T.: Bi-temporal Data – Theory and Practice. Morgan Kaufmann, San Francisco


  98 M. Kvet and K. Matiasko


8. Karima, T., Abdellatif, A., Ounalli, H.: Data mining based fragmentation technique for

distributed data warehouses environment using predicate construction technique. In: The 6th International Conference on Networked Computing and Advanced Information Management (2010)


9. Kvet, M., Matiaško, K.: Transaction management. In: 9th Iberian Conference on Information

Systems and Technologies (CISTI) 2014, Barcelona (2014)


10. Kvet, M., Matiaško, K.: Temporal transaction integrity constraints management. Cluster

Comput. 20(1), 673 –688 (2017)


11. Pedrozo, W., Vaz, M.: A tool for automatic index selection in database management

systems. In: 2014 International Symposium on Computer, Consumer and Control (2014)


12. Tuzhilin, A.: Using Temporal Logic and Datalog to Query Databases Evolving in Time.

  Forgotten Books, London (2017)


Acoustic Signal Analysis for Use

in Compressed Sensing Application

1(&) 1 1 Veronika Olešnaníková , Ondrej Karpiš , Lukáš Čechovič , 2 1 and Judith Molka-Danielsen

Department of Technical Cybernetics,

Faculty of Management Science and Informatics, University of Žilina,


Žilina, Slovakia


Faculty of Logistics, Molde University College, Molde, Norway

Abstract. The subject of the study was the possibility of application of com-

pressed sensing in wireless sensor networks. Compressed sensing is a

fast-forwarding area that provides the basis for data acquisition methods in

which it does not apply established signal sampling procedures but, subject to

certain conditions, it allows to significantly reduce the number of signal mea-

surements. As a result, it may be faster to obtain data or reduce energy demands

on data acquisition. This paper is dedicated to the investigation of suitable types

of acoustic transmitted signal for localization purposes in Wireless Sensor


  Keywords: Wireless sensor network Compressed sensing Localization application Acoustic signals

1 Introduction

  Wireless Sensor Networks (WSN) are a relatively new technology in intelligent environments. Use of WSN brings possibility to monitor the required information from the surroundings. This system consists of spatially distributed autonomous modules that are capable of interacting with each other. Nodes are located in the observed area and continuously evaluate the status of the monitored objects. Required sensor may differ based on the application. An example of this can be the recording of acoustic emissions in traffic, monitoring of person’s movement in an object, the acquisition of meteorological data, or the wide use of WSN in intelligent buildings and most recently in various IoT applications. Interaction between nodes is provided by RF communi- cation. According to

   ], up to 99% of the installed sensors currently communicate by

  wires. The assumption for the next 10 years is that WSN technology should cover 10% of all sensor networks. The communication in the sensor network is often significantly limited. Individual modules need to be placed often in an areas without the presence of the power grid, resulting in a large energy limitation. Computational and transmitting power of the nodes has to meet this strict requirements. Gateways have defined per- meability depending on the application and the available amount of energy. Using WSN technology requires minimal installation and maintenance requirements.


V. Olešnaníková et al.

  For this reason, the use of compressed sensing (CS) in WSN is beneficial. Using CS, it is possible to considerably reduce the volume of data being transmitted, and at the same time it is possible to transfer all the data, not only the partial results, as is often the case of reducing the energy demands on the network.

2 Compressed Sensing

  The original idea of data compression was to collect all the data, execute the appro- priate transformation, evaluate the fundamental coefficients and discard most of the data, because for some reason they carry only a small amount of useful information. Compressed sensing comes with a different strategy: assuming the signal sparsity in some suitable domain straightly sample the signal only so many time as is really needed. The subsequent reconstruction of the signal leads to solving a system of undetermined linear equations with an infinite number of solutions. Among infinitely many solutions there can be found those which have the most of the unknowns equals to zero at the same time (the so-called sparse solutions) are required, because these solutions are usually very likely to be correct. This thematic area goes back to the earliest use of the Matching pursuit

   ] algorithm and the sparse decomposition of the

  signal to atoms using Basis pursuit

   ]. Compressed sensing was named as a research

  area in 2004 at Standford University [


  2.1 Sparse Representation of Signals The signal y (e.g. sound, image of video) can be represented as linear combination of the basis vectors a i X y x a

  ¼ i i ; i where x are the weights or coordinates y in the system a . Each a are called atoms and i i i the whole system is called dictionary. This model can be expressed as a system of linear equations in matrix form.

  Ax ¼ y; where x is the vector of unknown and y is the known vector. Matrix A is called also representative system.

  So k-sparse vector is such a vector, which has at the most k nonzero components. By the relative sparsity of the vector x of the length N we understand the ratio k/N. N Next, the C is the set of all k-sparse vectors with the length N.

  However, the real signals are not sparse the same as defined above. Instead of zero components, they contain small nonzero values. Therefore, it is also appropriate to define an approximation error.

  Acoustic Signal Analysis for Use 101

  2.2 Sparse Solutions of the Linear Systems Equations It is necessary to solve the classical system of linear equations Ax

  ¼ y, however, the unknown and sought vector x should be as sparse as possible, it means, that it should contain as many zero components as possible. The task is: min x subject to Ax m x k k ¼ y; where the vector y is known (observation, measurement, signal) and matrix m

2 C


  A . We suppose only cases where m\N ; m

  2 C N respectively, and A is the matrix with full line rank. Vector x contains only few nonzero components (coordi- nates), so observed vector y is the linear combination of only some columns of the matrix A.

  Each x, which meet the condition Ax = y, are called acceptable solutions. It is known from linear algebra that given the above condition, the matrix A permits end- lessly many solutions and forms affinity space. Matrix A guarantees the possibility of the reconstruction of the original x from y.

  2.3 Compressed Sensing Compressed sensing is probably one of the most attractive applications in the field of sparse signal representation. The main idea of compressed sensing (compressive sampling) lies in the non-adaptive sampling of signals only by the number of samples that is really needed. This is the same problem - searching a sparse solution to the task min x subject to Ax - minimization. 1 x k k ¼ y, with applying l

  The sparsity of the signal is assumed in given dictionary, DFT base in our case. Let the base be labeled ¼ W x, where x is k-sparse. W so the signal can be expressed as z

  The goal is to provide small number of nonadaptive measurements, that will have the character of scalar products with a signal, which can be expressed as y = Pz = P Wx. Where P is so called measuring matrix of size m

  N and each components of vector y are the results of measurement, which originate as linear combination of signal samples. By the term “sample” is meant scalar, which is created by the linear com- bination of original values of the signal. Number of measurement is m

  N. From the Fig.


it can be seen, that matrix A, called dictionary, has in this case form A = P

  W and generally it is required to solve the l -minimization task. 1 min x subject to y x k k ¼ PWx: The basic task is to find such a measurement matrix P, so it would be possible to reconstruct signal from little number of samples. Suitable measuring matrices should be in the form P = R

  U. Where U is the matrix N N and R is the matrix, which is created from the identity matrix N N by selecting only m randomly chosen rows. It means, it serves as random selection of rows from matrix

  U. Random selection R is in boundaries of normal distribution. Total form of measuring matrix A is than com- pound of matrices A = R

  captured UW. The whole scheme is illustrated in the Fig.


V. Olešnaníková et al.


Fig. 1. Illustration of the situation in compressed sensing (without noise).

  vector y is equal to product of measurement matrix P = RU, the identity matrix W and sparse vector x. In the process of sampling there is a vector z = Wx, which is not sparse by itself, but it is sparse in given basis W, in the picture in orthonormal basis of DCT. Matrix U has Gaussian distribution. Matrix R was created from identity matrix N

  N by keeping m rows. In the pseudo-colored scheme, the blue color represents zero and warmer color represents higher value ].

  In order to reconstruct the signal by the l 1

  • minimization, it is necessary to define how many measurements need to be done (count of rows m of the matrix P). In case of random measurements matrices R the number of measurements depends on mutual coherence (µ). In cases, the matrix is compound from two orthogonal basis

  U and W, so ½ Š is

  W; U


  w l ; / ð ½ W; U Š Þ ¼ min i j 1

  i;j N 1 p

  and the value of coherence l is inside the interval ffiffiffi and 1. N Next statement give the condition when the precise reconstruction from the mea- surement is guaranteed.



  Let’s have the signal z, which had in the basis W k-sparse representation x. Than the solution of l -minimization is 1 min x subject to y x k k ¼ RUWx; 1 where y are measurements, it is at the same time with high probability the most sparse, while the number of row matrix is: 2 m

  C l ð ½ Š Þ k N lnN; W; U for the specific constant C.

  It means, that number of measurement depends on sparsity of the signal only linearly. Coherence influences quadratic the necessary number of measurements. Therefore, it is also an effort to look for such a pairs, whose coherence is minimal - for

  Acoustic Signal Analysis for Use 103

  p ffiffiffiffi pairs N it is enough to provide k ½ W; U Š with coherence 1= lnN measurements.

  However, if the coherence increases measurement according this condition stops to be relevant, because the number of measurements m will exceeds the number of signal samples N (e.g. in case U ¼ W).

  2.4 Localization via WSN Compressed sensing methods can be effectively implemented in the task of localization of an object. Searching for the proper location of objects is no longer a matter of just tion halls or logistics centers, or combine both environments. There are two types of localization tasks. It can be an absolute location which take in the consideration geo- graphic coordinates. The second type is the relative location and the position of an object is given relatively to the environment in which the object is located.

  This localization task scenario is based on the use of the acoustic signal. The baseline task is considering the use of WSN modules which are equipped with a

  is speaker. These transmit nodes are labeled as V1–V3 (Fig.

  designed to record the transmitted (acoustic) signals. Module D1 is considered as the target we are supposed to localize. Applying of the CS to the recorded signal and sending the sparse samples for processing will be covered by the detection node D1. The centralized topology in counts with the idea that obtained data will be transmitted and processed by the central node C1 with sufficient performance parameters. There will be the reconstruction algorithms applied to the received samples and subsequently the position of the object can be evaluated. The principal scheme of the task is shown in Fig.



Fig. 2. Illustration of the simulation example.


V. Olešnaníková et al.

  Several terms are explained in the following, which are subsequently used in tables and text. Compression ratio (KP) Let N be the number of samples of the signal y(n) gained by uniform sampling with sampling frequency f given by the Shanon theorem in given interval and let m be the vz number of samples of the signal y(n) selected on the same interval based on the measurement matrix P, than the compression ratio is defined as

  N KP : m

  KP is a basic indicator if it is applicable to use a compressed sensing. The task is to maximize KP while maintaining the maximum permissible reconstruction error. Reconstruction error Error of the reconstruction is define as mean absolute deviation of the difference of signals X N

  1 e r n ¼ j ð Þ j;

  N n



  where the difference of the signals is r(n) = y(n) − y’(n), and y’(n) is reconstruction of the original generated signal y(n) from the samples gained by CS and N is the count of elements of the vector r(n). By multiplying with constant K = 100 we present the error in percent below. The frequency domain error is determined analogously to the mean absolute variance of the frequency spectrum difference.

  Other terms found in text and tables: 1 , f 2

  • f – carrier frequency of the generated signal, vz
  • f – sampling frequency, [Hz]
  • measurementmatrix – number of rows (m) of the measurement matrix P

  2.5 Transmitted Signal This task is considered using an acoustic signal. The basic task is to examine the type of broadcast signal - the signal can be transmitted continuously over time (continuous) or in batches over the time (burst). Continuous In the case of a continuous signal transmission, the reconstruction is less demanding, which means that the compression ratio can reach higher values at the desired devia- tion, see Table

  The resulting error can be considered in practice to be zero, but it is

  an ideal case (no noise) when the sampled signal is maximal. Further, it can also be seen from Fig.

   that the signal before and after reconstruction shows a minimal error.

  The disadvantage of this sampling, however, is that the localizing task is not trivially solvable and it is not clearly evident in which period the signal is located, so the determination of the distance is not unambiguous. It is possible to implement other

  Acoustic Signal Analysis for Use 105

Table 1. Results of the reconstruction with use of KP = 100.


Name Carrier f[Hz] f [Hz] SNR[dB] Error [%]


cont_1 100 2000 −6,1E-9 – cont_2 200 and 500 2000 1,2E-8 cont_1a 100 2000

  11 14,8 cont_2a 200 and 500 2000


  35 cont_1b 100 2000 2,5 10,28 cont_2b 200 and 500 2000 2,5

  25 Fig. 3. Signal cont_2 from the Table before and after reconstruction

  software solutions, e.g. inserting auxiliary signals into continuous broadcasting to ensure unambiguous timelines. However, with this solution, there are also increasing demands on the performance of individual nodes in the network, which is an unde- sirable effect.

  Burst By modifying the transmission from continuous to burst, the tests confirmed that the compression ratio would be lower in this case. However, it is possible to analyze more accurately the captured signal. Under ideal conditions, it is possible to determine the beginning of the broadcast quite accurately. The tests were performed using an acoustic burst signal in the range of 100 Hz to 1 kHz.

  Simulations have confirmed that reconstruction can be considered sufficient if the error between the original and the reconstructed signal is up to 5%. In reconstructed signals that contain one carrier frequency with a fault rate of up to 15%, this frequency is still strongly represented, with the time domain signal not being usable without


  further filtration. In Table are the selected experiments - in the first approach, the tested burst signals contained one carrier frequency. Each of these items repeatedly applied 100 tests, and the error column indicates the average error of all iterations. In the next table, Table

   the signal was generated using two different carrier frequencies, the frequency ratios being determined randomly. During the testing this type of broadcast several important facts have been observed. Upon reconstruction of the burst signal there is a high error rate, mainly at the beginning and end (Fig.

  of the burst signal. If the same compression ratio as in

  the continuous signal is kept, an unusable rate of reconstruction has been achieved as

Table 2. Tests of the basic burst signals and their parameters.

  Name Carrier f[Hz] f vz [Hz] Measurement matrix KP Error [%] burst1f_01 100 2000 200 10 3,73 burst1f_02 200 2000 200 10 3,95 burst1f_03 300 2100 250 8,5 2,11 burst1f_04 400 2000 200 10 3,26 burst1f_05 500 2000 200

  10 2,05 burst1f_06 800 4000 300 13 1,89 burst1f_07 1000 4000 200 20 1,6 burst1f_08 2000 8000 300 27 1,09

Table 3. Tests of the combination of burst signals and their parameters.

  Name f 1 [Hz] f 2 [Hz] Frequency ratio KP Measurement matrix Error [%] burst2f_01a 100 300

  3 13,33 300 3,78 burst2f_01b 100 400 4 13,33 300 1,5 burst2f_01c 100 130 1,3

  20 200 1,27 burst2f_02a 200 130 0,65 20 200 3,02 burst2f_02b 200 350 1,75 13,33 300 1,32 burst2f_03a 500 144 0,288 20 200 2,68 burst2f_03b 500 630 1,26 13,33 300 1,58 burst2f_04a 1000 1300 1,3 13,33 300 1,53 burst2f_04b 1000 840 0,84 13,33 300 1,17


V. Olešnaníková et al.

  Acoustic Signal Analysis for Use 107

Fig. 5. Compare of continuous and burst signals after reconstruction with KP = 200.

  can be seen at Fig.

   where the upper two pictures represents the burst signal and the

  lower two pictures the continuous signal. This results led to the next adjustments of the generated signal.

  The task is to generate a signal which representation in the frequency domain is as clear as possible. For the purpose of smoothing the spectrum the weighing windows were applied. Figure

   illustrates the use of the window functions and effect on the signal frequency spectrum. In most cases, real-world signals are close to sparse signals.


Fig. 6. Frequency spectrum after reconstruction with and without using the window function.


V. Olešnaníková et al.

  The displayed result is a test with Hamming window that confirmed improvements in re-presentation in the frequency domain. If the signal is not sparse in given domain, the reconstruction by the L1-minimization will gain more significant errors.

  3 Conclusion

  The possibilities of using WSN applications are constantly expanding. We have explored the possibilities of using compressed sensing in an object localization application in wireless sensor networks. Compressed sensing is based on the theory of a sparse representation of data. Therefore, the main task was to find a suitable signal that is sparse in a certain domain. Another signal requirement was to be easily generated with limited HW means that are often present in WSN nodes. An acoustic signal has been selected that fulfills all the requirements and therefore we have used it in a simulation example. When designing the transmitted signal, it was necessary to determine the carrier frequencies of the signal. With the first type of transmission, a very good compression ratio was achieved. However, further implementation of syn- chronization signals would be needed to analyze it, which would displace the original application of the compressed sensing aside. The batch method of transmitting the signal has been selected. Although we achieve a slightly lower compression ratio, we can accurately analyze it. However, this method of transmission also results in decreasing of signal sparsity in frequency domain. The signal contains except for transmitted period also a zero-value samples, and superimposed noise throughout the transmission time. Thanks to the use of weighting windows, the reconstruction error has been greatly reduced. The benefit is also definition of the type of broadcast as burst signals. The appropriate broadcast configuration is dependent on room size and also on the required accuracy.



1. Wireless sensor networks.

iDTechEx 2008. 30 August 2016


2. Mallat, S., Zhang, Z.: Matching pursuits with time-frequency dictionaries. IEEE Trans. Signal

Process. 41(12), 3397–3415 (1993).


3. Chen, S.S., Donoho, D.L., Saunders, M.A.: Atomic decomposition by basis pursuit. SIAM

Rev. 43(1), 129–159 (2001).


4. Donoho, D.: Stanford University. . 21 March


  5. Candes, E.: Stanford University. .

  21 March 2017

6. Candes, E.J., Wakin, M.B.: An introduction to compressive sampling. IEEE Signal Process.

  Mag. 25(2), 21–30 (2008). ISSN 1053-5888,

7. Hrbáček, R., Rajmic, P., et al.: Rídké reprezentace signálu: komprimované snímání.

Elektrorevue 13(6) (2011)

  Community and Public Collaboration


Applying Recommender Approaches

to the Real Estate e-Commerce Market

1(&) 1 2 1 Julian Knoll , Rainer Groß , Axel Schwanke , Bernhard Rinn , 1 1 and Martin Schreyer

Technische Hochschule Nürnberg,

Keßlerplatz 12, 90489 Nuremberg, Germany





Immowelt AG, Nordostpark 3-5, 90411 Nuremberg, Germany

Abstract. Like in many other branches of modern economies, the internet

changed the behavior of suppliers and customers in real estate markets. Rec-

ommender systems became more and more important in providing customers

with a fast and easy way to find suitable real estate items.


In this paper, we show different possibilities for embedding the recommen-

dation engine into the user journey of a real estate portal. Moreover, we present

how additional information regarding real estate items can be incorporated into

the recommendation process. Finally, we compare the recommendation quality

of the state-of-the-art approaches deep learning and factorization machines with

collaborative filtering (the currently used recommender algorithm) based on a

data set extracted from the productive system of the Immowelt Group. Keywords: Recommender system Real estate Deep learning Factorization machine

  Collaborative filtering

1 Introduction

  For both suppliers and customers, it is very difficult to get an overview of the real estate market. Real estate portals on the Internet can reach a large number of users with detailed offer information. Due to the amount of available information, the users of these portals are faced with the question of how to effectively and efficiently identify the real estate items that are relevant to their personal needs. Recommendation systems deal with this problem and try to identify the relevant items based on the collected data about users, items, and interactions between them. is one such real estate portal. This sector is characterized by the fact that there are a large number of new users every day. Many of them are occasional users who do not register and remain only for a few transactions. There are also many real estate items every day. In contrast to users, these have many more attributes and characteristics. However, items can only be brokered once before they are removed from the portal. They sometimes only exist for a few days or weeks. The Immowelt recommendation engine currently uses a collaborative filtering (CF) approach which

  112 J. Knoll et al.

  recommends items that are outside the user’s search criteria but have already been visited by similar users. In addition, for good recommendations the CF algorithm requires that new items are visited by a minimum number of users (cold start problem).

  In this paper, we examine the quality of the recommendation of two state-of-the-art techniques using data from the productive system of the Immowelt Group. Further- more, we investigate how the algorithms behave in cold start situations and real estate specific use cases.

  We make the following main contributions. First, we provide a structured repre- sentation of the user journey in a real estate portal and show where a recommendation regarding items into the recommendation process of Immowelt Group, which so far is only capable of including relations between users and items. Third, we compare the state-of-the-art approaches deep learning (DL) and factorization machines (FM) with the currently used collaborated filtering (CF) algorithm of the Immowelt Group.

  The remainder of this work is structured as follows. In Sect.

  we give an overview

  of the related work. The application of the recommender algorithms is presented in Sect.


2 Related Work

  2.1 Recommender Systems Within the Real Estate Domain One of the first approaches that outlined the application of recommendation systems on the real estate domain was described by Shearin and Lieberman

   ] in 2001. They

  developed a program that attempted to reproduce the results of a human real estate agent based on an interactive learning procedure which was not further specified. To obtain information about the user’s preferences, they used an extensive questionnaire that the user was required to fill out before receiving any suggestions. This information was then either used to set a standard filter on a database or to find real estate items depending on the interactive learning procedure.

  A different approach referring to geographic recommendations was published by de Graaff et al.

   ]. They describe a geosocial recommender system which includes many

  data sources and should be capable of recommending local businesses, especially based on data from social media. The approach applies profile matching, which is based on a collaborative filtering algorithm, to select relevant items. The authors claim that their concept of geoprofiles enables this approach to be transferred to the domain of real estate items without presenting further details.

  Yuan et al.


  ] aimed to improve the efficiency and affordability of an online housing search. They developed a user-oriented recommendation system for real estate websites using ontology and case-based reasoning and focused on user requirements and search behaviors. Their target was to find the semantic construction of housing unit information and all sub-mechanism designs for constructing a recommendation system. Based on a questionnaire and direct user observation, they tried to extract the semantic Applying Recommender Approaches to the Real Estate e-Commerce Market 113

  relationships between information segments into an ontological structure. The extracted ontology then contained information about the knowledge collected in the user study and real estate domains. Finally, they specified the case representation and case indices to calculate the similarity between the queries of a user and cases.

  Another simulation study based on real estate data in the field of recommender systems was published by Chulyadyo and Leray [

  who proposed a recommender

  approach to build a personalized Probabilistic Relational Model (PRM) based on users’ preferences on decision making criteria. They showed that content-based, collaborative filtering and hybrid approaches can be achieved from the same PRM to increase the

  In the real estate sector, there have been no publications to date on recommendation systems based on DL or FM. Therefore, in the following sections we have a general look at the applications of these approaches in the recommender area and apply them to the real estate sector.

  2.2 Deep Learning as a Recommender Algorithm In recent years DL has become increasingly popular. Numerous new approaches based on neural networks have also been developed in the field of recommendation systems.

  As a result, the annual number of scientific publications on this topic has begun to grow exponentially. In their survey paper on deep learning based recommender systems, Zhang et al.

   ] analyzed the employed neural network models in literature. In this

  context, the term “neural network model” refers to the architecture of the artificial neural network used within the recommendation process.

  First of all, they distinguish between approaches that use several different types of neural networks and those that use only one type of neural network. Approaches that are based on only one type of neural network are further classified according to this network. The most common approaches in this context are Multilayer Perceptron (MLP), Autoencoder (AE), Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN). The intention of these approaches is outlined briefly in the following:

  • As a universal approximator, the strength of MLPs lies in the fact that they can approximate any measurable function to any desired extent

   ]. Recommender

  algorithms can profit from this by using an MLP to predict the interest of a user towards different items. The “Neural Collaborative Filtering” approach of He et al. [

  for example, employs an MLP to estimate the interaction of users and items. Its

  architecture consists of two input layers, one for the identification of users and one the specification of items. As a supervised learning algorithm, two different pieces of information are passed to the MLP during the training phase: (1) pairs of users and items and (2) a label that indicates whether this item is relevant for the specific user or not.

  • An AE neural network is an unsupervised learning algorithm, setting the target values to be equal to the inputs. In other words, it attempts to learn an approxi- mation to the identity function. One of the first publications concerning the gen- eration of recommendations with AEs was described in 2014 by Ouyang et al. [
  • CNNs are mainly used to process visual, textual or audible data. Accordingly, characteristics from such data. For example. Gong et al.

  • Approaches based on RNNs mainly use them to model the temporal dynamics and sequential character of user interaction with items. This can be useful, for example, if the user cannot be identified via a login but only on the basis of a session. In such cases, the data on the mostly implicit feedback is even thinner than usual. Approaches based on recurrent neural networks can lead to good results by including the sequential character of session data ].

  individual decisions in tweets. Since they had to deal with a pervasive cold-start problem, they extracted the interests of users based on rich information features, including textual content, employing so-called Co-FMs.

  ] employed seconder-order FMs for collaborative filtering based on both explicit and implicit feedback (ratings and mouse clicks). They set the compressed knowledge of user homophily and item correlation similarly and thus developed a two-step algorithm including the mining of compressed knowledge and the integration


  Pan et al.

  recommendations by transferring knowledge from an auxiliary domain (e.g., songs) into a target domain (e.g., videos). In a simulation study of this so-called cross-domain collaborative filtering, second-order FMs yielded the best results of the considered algorithms.

  analyzed how to improve

  Based on a data set from Amazon, Loni et al. [

   ] aimed to predict user interests and

  They used one AE network per user, but set the same number of hidden units for each of these AEs. Every AE only contained input units for the movies rated by the specific user. Consequently, an AE had few connections if the user rated few movies. Each hidden unit could then represent a significant dependency between the ratings of different movies. To include the information of other users, all of the corresponding weights and biases were tied together. Thus, if two users had rated the same movie, their two AEs used the same weights between the input/output units for that movie and the hidden units.

  Since this newly introduced FM approach was demonstrated within the recom- mender domain, it is hardly surprising that FMs often served in publications as rec- ommender algorithms. For example, Hong et al.

  context of recommender algorithms. In this introduction of the FM approach, he described the FM model and how to implement it with linear complexity. Furthermore, he conducted a simulation study which revealed that the FM approach is able to outperform support vector machines.

   ] based his description of the FM approach on an example in the

  2.3 Factorization Machines as a Recommender Algorithm Initially, Rendle

  so-called “attention-based CNN” for performing the hashtag recommendation task for microblogging services. They adopted the structure of CNNs in order to avoid the manual selection of features. To incorporate a trigger word mechanism, they propose the attention-based CNN architecture consisting of a local attention channel and global channel.

   ] proposed a novel

  114 J. Knoll et al.

  Applying Recommender Approaches to the Real Estate e-Commerce Market 115 ]

  By using a so-called field-aware FM for the RecSys 2015 Contest, Yan et al. achieved the third-best result in an e-commerce item recommendation problem. Their method was divided into different steps: map the top-N recommendation task to a binary classification problem, use second-order FMs and gradient boosting decision trees to obtain derived features, and build an ensemble of two second-order FM models trained on different feature sets to obtain recommendations.

  Another approach was described by Chu and Tsai [

  who proposed to include

  visual information in the data fed to the Factorization Machine. They advocate that, in addition to text information or metadata, restaurant attributes and user preference can the same space. Through experiments with various settings, they show that visual information can help to increase the quality of restaurant predictions.

3 Application of the Recommender Approach

  3.1 User Journey The Immowelt Group’s real estate portals record a total of over 50 million visits per month. Since Immowelt accompanies its customers throughout the complete journey, from searching to moving into their new home, recommendations are played out at different points of this journey. Figure


  shows a simplified view of the user journey of an Immowelt customer looking for a real estate item. Gray colored boxes illustrate the embedding of the recommendation engine within the journey.


Fig. 1. Immowelt user journey.

  116 J. Knoll et al.

  A typical user journey begins with a search for the desired real estate item by entering criteria such as location, real estate type, distribution type, price, living space, and number of rooms (1) (Fig.


  The search results in a list of suitable items (2) (Fig.


  Since the user has not visited an item yet, we do not know anything about his/her preferences and no alternative items can be recommended. The journey can also begin via the statistics page if the user wants to see statistical information of a certain city (4). From there it is possible to view various lists of real estate items from selected districts.

  (a) Search criteria (c) Recommendation list (b) Search result list


Fig. 2. search web pages.

  If the user is interested in a specific item, a click opens the corresponding item page with detailed information (3) where the user can scroll through the page and view all images. It is possible to memorize the item, make a note, or hide the item from further lists. With these actions, we learn about preferences of the user for the first time. Now the recommendation service can take effect and recommend items visited by other users with similar preferences.

  From now on recommendations can be added to every result-list (5) (Fig.


  c). If the user wants to see the complete list of recommendations, a link at the end of the list opens the recommendation page (6). If the user does not want to miss any suitable items, it is possible to create a search request (7) to be informed about new items by e-mail (8). If the user particularly likes a certain item, it is possible to contact the supplier to arrange a viewing appointment (9). Free space on the contact request confirmation page is used to show further recommendations. Contacting a supplier usually initiates the last phase of the item search which, in the event of a success, leads to a rental or purchase contract and ends with the move to the new home. Applying Recommender Approaches to the Real Estate e-Commerce Market 117

  3.2 Data Engineering The data examined during our simulation study is provided by Immowelt Group. It was extracted from their productive database with information about real estate items, users, and interactions between them. Besides user ID and item ID, the data set also provides some attributes for the real estate items (Table



Table 1. Original attributes.

Attributes Datatype Value range Description User ID Integer

  1–32,766 Uniquely identifies the user Item ID Integer 1–3,619 Uniquely identifies the item Real estate String Apartment or Differentiates between house and type house apartment Distribution String Rent or purchase Differentiates between rent and type purchase Purchase price Integer Purchase price in Euros

  0–5.4 mio. Basic rent Integer Rental price in Euros 0–4,500 Living space Integer Living space in square meters 0–1,380 Rooms Float Number of rooms 0–52

  We transform the attributes real estate type and distribution type from strings to indicator features (one hot encoding) in order to provide them to the recommender algorithms. The attributes purchase price, basic rent, living space and rooms have to be normalized, which is a standard procedure in many machine learning approaches. We use simple rescaling (min-max-scaling) to fit the values in a range from 0 to 1.

  Since there is no way for the users to give explicit feedback, a score indicating the relevance of the real estate item to the user has to be determined. We decide to pursue a common approach for implicit feedback by considering all reactions towards real estate items (clicks, views, filling in the contact form) as positive feedback and set their score to 1. Since we are only able to observe positive feedback, we need to generate negative feedback. Following Wu et al. [

  we produce negative feedback from the unob- served user item interactions and set five randomly sampled interactions to 0.

  3.3 Algorithms Collaborative Filtering. The currently used recommendation engine at Immowelt employs an item-based collaborative filtering (CF) approach. It matches the user’s visited real estate items to similar ones, and combines them into a recommendation list. In the first step, the similarities between all pairs of items need to be calculated. The resulting item-to-item similarity matrix is built by iterating through all user item interactions within a certain period of time (last 30 days), counting all item visits C i

  ð Þ n as well as all combinations of items visited together C i ð n ; i m Þ, and computing the adapted Jaccard similarity for each item pair using Eq.


  118 J. Knoll et al.

  2 C i ð ; i Þ n m sim i ð n ; i m Þ ¼ ð1Þ C i ð Þ þ C i n ð Þ m

  Afterwards, Recommendations are generated based on the similarity matrix and the list of items the active user has already visited. The algorithm selects all similar items and recursively aggregates them (Fig.

   ). After removing previously visited items, the system recommends the most similar ones to the active user. parent_sim=1.0; recursion_level=2; min_sim=0.001; rec_list=[] AddItem(recursion_level, item_list, parent_sim):

  If recursion_level > 0 For each item of item_list not already in rec_list Get sim_list of tuples ( ) from item-to-item matrix for For all of sim_list not already in rec_list weighted_sim = parent_sim if weighted_sim min_sim

  Add item-sim tuple ( ,weighted_sim) to rec_list AddItem(recursion_level-1, , weighted_sim) Get item_list of items visited by the user AddItem(recursion_level, item_list, parent_sim) Remove already visited items from rec_list

Fig. 3. Calculation of recommendations.

  Deep Learning. One of the state-of-the-art approaches we choose to compare with the currently used recommender algorithm is deep learning (DL). Since this is a basic DL approach that is easy to understand, we employ an MLP architecture based on the afore-mentioned Neural Collaborative Filtering

   ] which we enriched by incorporating

  the possibility of including additional information about the items. This extension allows the model to profit from data about the item attributes (e.g., living space). Like the original approach, our architecture makes use of embedding layers, which aim to reduce the dimensionality of sparse features and thus save runtime. Figure

   shows the

  architecture of the neural network in more detail: • The input data consist of the user ID, item ID, and the prepared item attributes.

  • On the left-hand side, we prepare the vector which is fed to the wide layer in four steps: (W1) the user ID is passed to an embedding layer, (W2) the item ID is passed to an embedding layer, (W3) the attribute date is passed to a feed-forward layer, (W4) the results from (W2) and (W3) are concatenated, and finally (W5) the results from step (W1) and (W4) are multiplied element-wise.
  • On the right-hand side, we generate a vector which is passed to an MLP with n

  (deep) layers in a three-step procedure: (D1) the dimensions of user ID are reduced by an embedding-layer, (D2) the item ID is passed to an embedding layer as well, (D3) the attribute data is passed to a feed-forward layer, and (D4) the resulting layers are concatenated and fed to the first layer of the MLP.

  Applying Recommender Approaches to the Real Estate e-Commerce Market 119 Loss

  Score Target Func Ɵon



Wide layer Deep layer

… mul Ɵply ement-wise

  Deep layer Deep layer




Embedding Layer Embedding Layer Feed-forward Layer

.4 .1

0 0 1 0 0 0 0 … 0 0 1 0 0 0 0 … 0 0 1 0 …

user item a


Fig. 4. Neural net architecture of the deep learning approach following He et al. [

  • To calculate the target value, the wide layer and the last deep layer are concatenated

  (A1) and connected to the output neuron. This neuron delivers a score expressing the preference of the active user towards the item specified in the input data.

  In our simulation study, the neural network is trained with a learning rate of 0.001 and a batch size of 1024. These parameters were determined in preliminary studies. In regards to the architecture, we use 264 neurons for the embedding layer W1/D1 and 256 neurons for the embedding layer W2/D2, while the deep layers of our 3-layer MLP had 256, 128, and 64 neurons. Factorization Machines. The approach of the factorization machine (FM) is a machine learning algorithm that can quickly process large amounts of data. FMs are linear regression models enhanced with interactions between two or more of the explanatory variables (features) whose coefficients are to be estimated on the basis of training data. This model can be used to predict or classify new patterns [ ].

  As a framework for the application of higher-order FMs, we use the R package FactoRizationMachines from Knoll


  ]. It is capable of estimating the coefficients in linear time by tremendously reducing the number of calculations necessary to estimate model parameters. In order to estimate the coefficients, the training data must be converted into a feature matrix X and a target vector y as shown in Fig.

   . Each row in

  the feature matrix X represents a prediction problem and possesses a corresponding

  120 J. Knoll et al. 1 1 1

  1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Fig. 5. Example of a feature matrix X and the corresponding target vector y.

  value in the target vector y. Each column in X represents an individual feature, which can either specify a user u j , an item i n , or an attribute a m (referring to a user, an item, or an interaction between them). The target vector y contains the corresponding scores, in our case 1 (positive feedback) referring to a relevant item or 0 (negative feedback) implying an irrelevant item.

  Higher-order FMs consider interactions between more than two features. In order to find suitable parameters for our simulation study, we conducted a grid search as part of a preliminary study. Based on the training data, various hyperparameter constellations were tested. The optimized FM model works with 13 second-order factors and 4 third-order factors.

4 Simulation Study

  4.1 Setup The goal of our simulation study is to evaluate the recommendation quality of DL and FMs in the real estate context and compare them with CF and the “global popularity” (GP) algorithm serving as a baseline approach. GP is a non-personalized approach based on the average score of the respective item over all users.

  Our expectation is that DL and FM should outperform GP and reach the level of CF. It should be noted that the data serving for the simulation study was generated while the CF approach was in use at Immowelt Group. Items of the ImmoweltData have been clicked and considered relevant in the simulation study because they have already been recommended by the CF algorithm in the past. If such an item is con- sidered as relevant in the test data set, the CF algorithm would rank this item higher since the algorithm has already reacted positively to this item in the past. Therefore, our simulation study shows too positive results for the CF approach. This bias must be taken into account when interpreting the results. We use 5-fold cross validation to ensure valid results. Figure shows the basic setup.

  In a subsequent analysis, we examine some of the results in more detail: On the one hand, we analyze the differences in the quality of recommendations between rental items and items for sale. On the other hand, we compare the performance of the approaches in a cold-start situation and in a non-cold start situation.

  Applying Recommender Approaches to the Real Estate e-Commerce Market 121 5-fold cross valida Ɵon


  Training data Model Model 28.7 k users

  ImmoweltData 32.7 k users

  Test data Applica Applica Applica Ɵon Ɵon Ɵon

  4 k users Quality of


Recommenda Ɵons: Ranking list Ranking list Ranking list Ranking list

AUC, Recall

Fig. 6. Basic setup of the simulation study.

  4.2 Metrics In an article about metrics used to evaluate recommender algorithms, Herlocker et al.


] differentiate between two major classes of metrics for different purposes. On the

  one hand, there are metrics for evaluating the accuracy of classifications, which usually consist of ratios of correct and incorrect decisions about the item recommendations. On the other hand, metrics for evaluating prediction accuracy measure how close the predicted ratings from the recommender system are to the true user ratings. Due to the metrics measuring the accuracy of predicted rating being less appropriate for evaluating the task of finding “good items”, in this article we focus on measuring the correct classification.


  To explain the intention of evaluating classification accuracy, we refer to Table in which a total of h items are categorized depending on their actual relevance to a user (irrelevant/relevant) and on whether they have been recommended by the recommender approach (not recommended/recommended).


Table 2. Categorization of items based on their relevance.

  Not recommend Recommended Total Irrelevant h h h 00 01 Relevant h h h 10 11 1 Total h h 1 h

  Cremonesi et al. [

  described an approach for evaluating recommender algo-

  rithms based on classification accuracy metrics. In this context, the decisive question is which items to consider as being relevant to a user and which not to. Therefore, the mentioned approach considered an item to be relevant for a specific user if this user assigned it with the highest rating, whereas non-rated items were considered as being irrelevant.

  122 J. Knoll et al.

  In line with this approach, we consider one real estate item with a positive feedback as being relevant for the active user, whereas we sample 499 items without positive feedback to generate the irrelevant items. In this article, we evaluate the different approaches with the two classification accuracy metrics “recall” and “area under curve”:

  • Recall: The recall metric is also known as “true positive rate” and looks at the proportion of relevant and recommended items h among all relevant items h . To
  • 11 1 calculate this proportion for a ranking list of suggested items, the parameter N, which determines the top-N items of the list considered as being recommended, has to be chosen. For instance, if N was set equal to 1 (Recall@1), only items rec- ommended at the first rank would be considered as recommended. Consequently, if the relevant item was recommended at the first rank in 5 out of 100 cases, the recall value would equal 5

      =100 ¼ 0:05. Recall metrics produce values between 0 (no relevant item recommended) and 1 (all relevant items recommended).

    • Area under curve: In the context of classification metrics, the ROC curve is a general approach which represents the graphical trade-off between the true positive rate (the recall) on the y-axis and the false positive rate (the percentage of recom- mended items among the irrelevant ones h =h ) on the x-axis. Consequently, an
    • 01 ROC curve can also be generated based on a ranking list suggested by a top-N recommender algorithm [

        To summarize the graphical ROC curve into one key

        figure, we calculate a value called “area under curve” (AUC). The better the rec- ommendation quality of the recommender approach, the higher the AUC value. It represents the probability that a relevant item is actually classified as such. The AUC metric is usually in a range between 0.5 (relevant items randomly assigned) and 1 (all relevant on the first rank) [

      5 Results

        The overall results of the simulation study described in the Section before are presented in Table


        The colored table fields show the column maximum for DL, FM, and GP. Since the results of CF are biased, values are displayed in gray in the last line of the respective comparison table and will not be interpreted in the first instance.

        First of all, we find that DL produces the best AUC and recall values, followed closely by the FM approach. Both approaches perform clearly better than the baseline approach “global popularity” (GP): The probability that an item relevant for the user will be recognized as relevant by the algorithm is 93.21% for DL, 91.94% for FMs, and 78.24%


      Table 3. Overall results.

        AUC @1 @5 @10 @25 @50 @100 @250 DL









        97.71 FM









        96.81 GP









        88.36 for GP. The probability with which an item relevant for the user appears in the first ten places of the ranking list is 51.07% for DL, 46.35% for FM, and only 13.26% for GP.

        Since the personalized algorithms (DL, FM) deliver much better results, the non-personalized baseline approach (GP) does not seem to be a meaningful alternative. DL and FM lead to results as good as the CF approach. Under consideration of the biased results of CF, DL and FM are at least sufficient. The corresponding ROC curves depicted in Fig.

        91.24 FM





        0.00 0.00 0.00 0.00


        90.18 GP

        0.00 4.81 92.67 10.80 33.20 47.23




        4.28 47.06 93.64 14.07 38.20 51.96

         are in line with these findings.






      Table 4. Results regarding the cold-start of “new items”.

      item cold-start no item cold-start AUC @1 @5 @10 @100 AUC @1 @5 @10 @100

        the best AUC and recall values, but at a relatively low level. The FMs deliver an even lower recommendation quality and the performance of GP is far behind. It seems that 0.0 0.2 0.4 0.6 0.8 1.0 0. 0. 2 0. 4 0. 6 0. TP 8 FPR R CF (92.95%) FM (91.93%) DL (93.21%) GP (78.18%)

      Fig. 7. ROC curves representing the recommendation quality of the examined approaches.

         left-hand side), DL again shows

        Referring to the cold-start of “new items” (Table

        context, we consider cold-start situations regarding “new items” (less than ten clicks) and “new users” (less than five positive feedbacks).

         illustrate the results of the analysis of the cold-start scenario. In this

        5.1 Cold-Start Performance Tables

        59.74 Applying Recommender Approaches to the Real Estate e-Commerce Market 123

        124 J. Knoll et al.


      Table 5. Results regarding the cold-start of “new users”.

      user cold-start no user cold-start AUC @1 @5 @10 @100 AUC @1 @5 @10 @100

        DL 91.87 13.22 35.36 47.78 87.68 94.91 14.56 40.29 55.26

        93.89 FM 90.73 10.88 31.01 43.47 86.12 93.47 10.25 34.58 50.01

        91.71 GP











        59.37 CF 91.64 15.50 41.29 54.26 89.86 94.62 10.20 33.72 49.74


        the customization of DL and FMs in this simulation study is not capable of dealing with the “new items” cold-start problem.

        In comparison to the situation with “new items”, the results for the cold-start of

         left-hand side) are notably better for all approaches and hardly

        “new users” (Table differ from the results of situations without cold-starts (Table

         right-hand side). In this case, the results of DL and the FM approach do not show remarkable differences.

        5.2 Domain Specific Analysis Since this paper is about applying recommender approaches to the real estate domain, it


        makes sense to analyze the differences regarding domain specific conditions. Table shows the comparison of the recommendation quality between houses and apartments. Like in the general analysis, DL reveals the best results for the house and apartment items, followed by the FM approach at almost the same level, and the GP algorithm far behind. However, if we compare the results between house and apartment items, it is remarkable that both show very similar results given the fact that they are recom- mended under very different conditions (e.g., house items stay much longer in the pool of recommendable items).


      Table 6. Comparison between houses and apartments.

      houses apartments AUC @1 @5 @10 @100 AUC @1 @5 @10 @100

        DL 94.88 18.74 47.63 61.92 93.48 92.87 12.81 35.48 48.87

        89.79 FM 92.32 16.13 40.81 53.68 89.80 91.86 9.48 30.91 44.86

        88.33 GP











        58.17 CF 92.64 19.04 49.17 61.80 91.93 93.01 11.98 35.69 50.34

        92.12 Table presents the comparison of the recommendation quality between rental and

        purchase items. Once again, the results do not differ very much and DL shows the best

        Applying Recommender Approaches to the Real Estate e-Commerce Market 125

      Table 7. Comparison between rental and purchase items.

      rental items purchase items AUC @1 @5 @10 @100 AUC @1 @5 @10 @100

        DL 93.78 14.72 39.14 52.58 91.19 92.20 12.21 34.72 48.42

        89.05 FM 92.69 11.15 34.19 48.11



        90.62 9.63 29.77 43.27

        85.78 GP

        82.36 2.89 10.53 16.97







        43.47 CF 93.30 13.10 38.10 52.90 92.19 92.33 13.29 37.72 51.17


      6 Conclusion and Future Work

        In this paper, we demonstrated where a recommendation engine can be embedded in the user journey of a real estate portal. Furthermore, we showed for the DL and the FM approach how additional information regarding real estate items can be incorporated into the recommendation process. Our simulation study, based on a data set extracted from the productive system of the Immowelt Group, delivered three major results: First, DL and the FM algorithm outperform the non-personalized GP approach and bring results at the level of the algorithm currently in use. DL und FM seem to be capable of dealing with the user cold-start problem, whereas our customization of these approa- ches was not able to cope with item cold-starts. Third, domain specific aspects (houses/apartments, rental items/purchase items) do not reveal vast differences regarding recommendation quality.

        Future work could be dedicated to examining model updates in real-time with respect to new items or new users. Furthermore, an analysis of whether the better recommendation quality of DL in comparison with the FM approach justifies the higher effort for training the DL model could be carried out.



      1. Shearin, S., Lieberman, H.: Intelligent profiling by example. In: Proceedings of the 6th

      International Conference on Intelligent User Interfaces, pp. 145–151 (2001)

      2. de Graaff, V., van Keulen, M., de By, R.: Towards geosocial recommender systems. In:

      Proceedings of the 4th International Workshop on Web Intelligence, pp. 8–12 (2012)


      3. Yuan, X., Lee, J., Kim, S., Kim, Y.: Toward a user-oriented recommendation system for real

      estate websites. Inf. Syst. 38(2), 231–243 (2013)


      4. Chulyadyo, R., Leraya, P.: A personalized recommender system from probabilistic relational

      model and users’ preferences. Procedia Comput. Sci. 35, 1063–1072 (2014)


      5. Zhang, S., Yao, L., Sun, A.: Deep Learning based Recommender System: A Survey and

      New Perspectives. arXiv preprint: (2017)


      6. Hornik, K., Stinchcombe, M., White, H.: Multilayer feedforward networks are universal

      approximators. Neural Netw. 2(5), 359–366 (1989)

      7. He, X., Liao, L., Zhang, H., Nie, L., Hu, X., Chua, T.: Neural collaborative filtering. In:

        126 J. Knoll et al.


      8. Ouyang, Y., Liu, W., Rong, W., Xiong, Z.: Autoencoder-based collaborative filtering. In:

      Proceedings of the 21st International Conference on Neural Information Processing, pp. 284–291 (2014)


      9. Gong, Y., Zhang, Q.: Hashtag recommendation using attention-based convolutional neural

      network. In: Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence, pp. 2782–2788 (2016)


      10. Rendle, S.: Factorization machines. In: Proceedings of the 10th International Conference on

      Data Mining (2010)


      11. Hong, L., Doumith, A., Davison, B.: Co-factorization machines: modeling user interests and

      predicting individual decisions in Twitter. In: Proceedings of the 6th ACM International Conference on Web Search and Data Mining, pp. 557–566 (2013)


      12. Loni, B., Shi, Y., Larson, M., Hanjalic, A.: Cross-domain collaborative filtering with

      factorization machines. In: de Rijke, M., Kenter, T., de Vries, A.P., Zhai, C., de Jong, F., Radinsky, K., Hofmann, K. (eds.) ECIR 2014. LNCS, vol. 8416, pp. 656–661. Springer, Cham (2014).


      13. Pan, W., Liu, Z., Ming, Z., Zhong, H., Wang, X., Xu, C.: Compressed knowledge transfer

      via Factorization Machine for heterogeneous collaborative recommendation. Knowl.-Based Syst. 85, 234–244 (2015)


      14. Yan, P., Zhou, X., Duan, Y.: E-commerce item recommendation based on field-aware

      Factorization Machine. In: Proceedings of the 2015 International ACM Recommender Systems Challenge, pp. 1–4 (2015)


      15. Chu, W.T., Tsai, Y.L.: A hybrid recommendation system considering visual information for

      predicting favorite restaurants. World Wide Web 20(6), 1313–1331 (2017)


      16. Wu, Y., DuBois, C., Zheng, A., Ester, M.: Collaborative denoising auto-encoders for top-n

      recommender systems. In: Proceedings of the Ninth ACM International Conference on Web Search and Data Mining, pp. 153–162 (2016)


      17. Knoll, J., Köckritz, D., Groß, R.: Markov random walk vs. higher-order factorization

      machines: a comparison of state-of-the-art recommender algorithms. In: Eichler, G., Erfurth,

      C., Fahrnberger, G. (eds.) I4CS 2017. CCIS, vol. 717, pp. 87–103. Springer, Cham (2017).



      18. Knoll, J.: Recommending with higher-order factorization machines. In: Bramer, M., Petridis,

      M. (eds.) Research and Development in Intelligent Systems XXXIII, pp. 103–116. Springer, Cham (2016).


      19. Herlocker, J., Konstan, J., Terveen, L., Riedl, J.: Evaluating collaborative filtering

      recommender systems. ACM Trans. Inf. Syst. 22(1), 5–53 (2004)


      20. Cremonesi, P., Koren, Y., Turrin, R.: Performance of recommender algorithms on top-N

      recommendation tasks. In: Proceedings of the 4th ACM Conference on Recommender Systems, pp. 39–46 (2010)


      21. Grottke, M., Knoll, J., Groß, R.: How the distribution of the number of items rated per user

      influences the quality of recommendations. In: Proceedings of the 15th International Conference on Innovations for Community Services, pp. 1–8 (2015)


      A Next Generation Chatbot-Framework

      for the Public Administration

      ( )


        Andreas Lommatzsch


      Agent Technologies in Business Applications and Telecommunication Group (AOT),

      Technische Universit¨ at Berlin, Ernst-Reuter-Platz 7, 10587 Berlin, Germany


        With the growing importance of dialog system and personal

      assistance systems (e.g. Google Now or Amazon Alexa) chatbots

      arrive more and more in the focus of interest. Current chatbots are typi-

      cally tailored for specific scenarios and rather simple questions and com-

      mands. These systems cannot readily handle application domains char-

      acterized by a large number of relevant facts and complex services (e.g.,

      offered by the public administration).


      In this work, we present a chatbot framework tailored specifically to

      the needs of public administrations able to provide answers for all types

      of questions related to offered services and offices. The challenges in

      this scenario are the large number of relevant services, the complexity

      of administrative services, the context-dependent relevance of user ques-

      tions, the differences in expert-language and user-language as well as the

      necessity of providing highly reliable answers for all questions.


      We present the developed framework and discuss our experiences

      obtained while running the chatbot publicly with real users. The results

      show that our system efficiently provides relevant answers to the user

      questions. We explain how our system extends existing approaches and

      discuss the experiences with the live system.

        Keywords: Chatbots Information retrieval Context

      · ·

        Human-computer interaction Natural language processing ·

        Question answering

      1 Introduction

        With the rapid advances in the processing of natural language input and the development of advanced machine learning methods chatbots (“artificial con- versational agents”) have entered into a variety of domains. Chatbots typically provide answers to frequently asked user questions based on a knowledge base. This reduces the load for human experts who do not have to answer the same questions again and again. Compared with human experts, artificial chatbots are always friendly, flexibly scalable, and ensure a constant level of answer quality. Furthermore, chatbots may support different natural speech-based input chan- nels in order to fit best with the user preferences.


      A. Lommatzsch

        Internet search engines (e.g. Google) represent an alternative source for information. Still, they do not provide concrete answers but merely return a potentially long list of documents. The disadvantage of this approach is that users must cope with an information overload requiring a lot of time in searching the desired information in the list of (often complex) documents. On the other hand, chatbots are designed to mimic human expert behavior. Their task is to concisely answer questions. This alleviates users from reading through bulks of large documents.

        Most existing chat bots are tailored for answering simple questions (“how is the weather forecast for tomorrow”). They fail answering more complex informa- tion demands. In this work, we introduce a framework that deals with complex questions regarding public administration services typically requiring a dialog with several interactions. Scenario Description.

        Citizens must use public administrations’ services to obtain official documents (e.g. passports) or register for certain programs (e.g. cash housing assistance). These services require following a specific pro- cedure involving administrative acts. As a result, many service-related questions emerge. The national service telephone number 115 has been set up in Germany to assist citizens with their information needs. The administration has decided to complement call center agents with a chatbot to reduce their workload and facilitate a 24/7 availability.

        Requirements and Research Questions.

        The objective of our work is the develop- ment of a chatbot optimized for answering questions related to the administrative services. The system must be able to identify requested services and locations as well as to provide the relevant information related to the specific properties of the requested services. In addition, the answers must be correct since user want to rely on provided answers. Furthermore, the chatbot should imitate a human call center agent; thus the system must support typical chat etiquette and small talk questions.

        From a research perspective, detecting the intent of a question represents a major challenge. The system must consider the context and detect whether user input should be handled as follow-up question or whether the user wants to switch to a new topic. If a question is related to a topic not part of the system, the chatbot must handle this and should provide recommendations how to get answers for these types of questions.

        Structure of the Work. The remaining work is structured as follows. In Sect.


        present the analyzed scenario in detail. Existing research related to our scenario is discussed in Sect.


        In addition we discuss the developed methods and the strengths of our solution. The results of the evaluation are discussed in Sect.

         A Next Generation Chatbot-Framework for the Public Administration 129

      2 Problem Description

        In this section, we present the analyzed scenario in detail and explain the specific requirements. We discuss the data sources and relevant APIs.

        2.1 Requirements Chatbots are typically designed to act as experts providing answers to fre- quently asked questions for a narrowly defined domain. Thus, bots should act like should be able to provide answers optimized to the specific user needs considering the questions’ context (e.g., the previous conversions and time). Furthermore, agents should proactively guide users to ask the right questions and give hints leading to useful information. In many scenarios, the agent must ensure that the answers are correct so that users can rely on the provided answers. This means that the agent must formulate the answers in a way that misunderstandings are avoided. If the agent cannot understand the meaning of the question, the agent should inquire the user’s information needs.

        2.2 The Analyzed Scenario in Detail In this work, we present a conversational bot used by Berlin’s administrative services and citizen offices (“B¨ urger¨ amter”). The system adds another commu- nication channel amid Berlin’s administration and its citizens. The chatbot’s duties include providing information on where citizen offices are located, their opening hours, which services they offer, which fees they charge, which pre- requisites must be met for individual services, which forms are required, who is responsible for services, services’ legal details, and more. Furthermore, the bot should refer to external information portals if services are offered by other administrative units (e.g., the federal government administration).

        2.3 The Data Basis Our system’s knowledgebase encompasses contents of Berlin’s administration describing services. The data are available in a tree-like structure as JSON-files.

        The data contain a list of services. For each service, meta-data are available either as plain text or HTML. Figure

         shows an exemplary service description.

        In addition to the service descriptions (provided by the Berlin’s adminis- tration), external knowledge sources are integrated such as the National Ser-


        vice Description Catalogue (“Leika. This catalogue provides synonyms and keywords for the services mapping formal “administrative” language to more colloquial terms. 1


      A. Lommatzsch

        Fig. 1.

        An service description, rendered as HTML (left), automatically translated into English (middle), and as JSON (right).

        2.4 Challenges Developing a chatbot for the scenario explained above, leads to several differ- ent challenges: First, the chatbot must provide answers to a large collection of complex services (≈700), locations (≈600), and “topics” (≈1500). The answers must be highly reliable since the chatbot acts as an official representative of the administration. Since the services are typically complex, the agent should know the different relevant aspects of the services, consider the context in user questions as well as guide the user through complex topics. Furthermore, the ser- vice descriptions are written in a formal administrative-specific style requiring a “translation” to the user language.

        Besides, the chatbot should mimic the behavior of human agents. The chatbot must remain polite and respond to “small talk.” If the bot fails to understand the information need, it should advise users how to change the question to get to the desired information (e.g. by suggesting improved questions).

        2.5 Discussion Chatbots represent an expanding technology helping users to satisfy their infor- mation needs in a larger number of domains. The development of systems is chal- lenging due to the integration of large knowledge bases, the necessity of managing complex dialogue considering the context and the individual user preferences. In addition, the system must be open to new knowledge sources and able to handle new trends.

        In the following section, we review existing frameworks. Subsequently, we introduce our chatbot framework and compare the features of our system with the strengths and weaknesses of existing systems. We discuss our experiences obtained running the chatbot system.

      3 Related Work In this section, we review existing approaches and systems.

        A Next Generation Chatbot-Framework for the Public Administration 131

        3.1 Personal Assistance Bots Conversation agents (“Chatbot”) got in the focus of interest in recent years. This can be explained by ubiquity of mobile devices and installed personal assis- tants informing users about the weather, calendar entries and allowing users to control smart home devices

         ]. For these use cases chatbot frameworks such



        as Amazon Lex (“Alex have been developed focusing on handling a rather small set of questions. Users can train their per- sonal chatbot so that it learns the user wording and adapts to the individual user preferences. Overall, these frameworks are only suitable for narrow domains characterized since a sufficient number of training examples for all user intents must be provided. In e-government scenarios these frameworks are not appro- priate due to the very large number of complex services that must be considered by the chatbot.

        3.2 IR Based Question-Answering Systems Chatbots are usually designed for answering questions applying a combination of different Artificial Intelligence approaches

        . Traditionally, Question-

        Answering systems are based on the mechanism of web forums. Users ask questions and other (human) users answer the question. Well-known systems are StackQverflow, Quora, or If a user asks a question, the Q&A systems typically compute the most similar already answered ques- tions. For the efficient search classic IR methods are used based on an inverted index representing user questions and past question-answering pairs in term vec- tors

         ]. The challenge with this approach is determining suitable similarity mea-

        sures and term weights. The tf-idf scheme (used by classic search engines) does not perform well since frequent words (such as interrogative pronoun) essential for understanding the question, are assigned with a very low tf-idf weights .

        In e-government scenarios, the IR-based approach is not directly applicable, but should be adapted to handle differences in the user language and the answers’ wording. In addition, the correctness and up-to-dateness of the answers must be ensured. Thus, answers should be generated based on a trusted, well-maintained knowledge base. Moreover, chatbots should provide concrete answers and not complex documents.

        3.3 Existing e-Government Chatbots Providing answers for government-related questions is an important domain for conversational agents

         ]. Several chatbots have been introduced to the e-

        government scenarios

        . Singapore’s administration operates a bot providing

        citizens with links to their web portal

         ]. The bot can answer a broad spectrum

        of questions but looks more like a traditional search engine than a bot imitating a human expert. 2 3



      A. Lommatzsch

        The “WienBot” integrates into Facebook Messenger

        . In contrast to Sin- gapore’s bot, it mimics humans as it provides short answers and uses emoticons.

        The bot seems to use a rather small knowledge base and a limited amount of ser- vice meta-data (synonyms). Overall, the system has good small talk capabilities but only a limited answer precision in the core domain (citizen services).

        In Poland, several cities use an e-government chatbot developed by inteli- . The system provides an avatar mimicking an expert for citizen ser- vices related questions. The system supports small talk and is able to correct typos. In our tests the system provided useful answers for about 2/3 of our ques- tions. For many questions it did not deliver a direct answer but explained who to contact in order to get an answer.

        There are several e-Government initiatives motivated by the e-Government law passed by the parliament in 2013

         ]. But still, only a few e-Government

        chatbots went public like the “Bundesadler” (federal eagle) tailored for answering question related to the German parliament. This system supports small talk well, but has limited capabilities with respect to handle natural conversions: It does not consider the dialog context and cannot answer complex questions requiring the discussion of several different aspects.

        3.4 Discussion Our analysis of existing work shows that chatbots combine several different meth- ods including, information retrieval and natural language understanding. Cur- rent systems can merely handle a limited number of defined questions. They tend to fail imitating human experts. Most existing systems do not support complex dialogs but focusing on returning a sentence or a link. Asking for more details or unclear aspects is not supported.

        These shortcomings of existing systems motivate us to advance the state of the art with an improved framework which supports large knowledge bases, handles a wide variety of small talk, and learns to map questions and answers. In addition a chatbot should support longer dialogues helping the user interactively finding all aspects relevant to a specific question.

      4 Approach In this section, we explain our approach and discuss the system’s architecture.

        4.1 Overall System Architecture We developed a hybrid approach generating answers based on FAQ databases, small talk collections and large structured knowledge bases. In addition, the system includes components for handling specialized requests such as questions in foreign languages, detecting sentiments, and small talk.


         visualizes the system architecture. A web application serves as the

        chatbot’s front-end. It handles http(s) communication and provides a REST

        A Next Generation Chatbot-Framework for the Public Administration 133

        API for alternative user interfaces. The query analyzer processes the users’ input. It classifies the information needs and forwards the request to the designated component. The intent classification considers the context and the current dia- log (session) ensuring that follow-up questions are handled correctly. Having predicted the user intention, the request is delegated to the best suited com- ponents. The user input is adapted and enriched with respect of the selected component incorporating several different knowledge sources (e.g. synonyms or geo-data). The selected components return the answer for the user question to the dialog management component. The aggregated answer is provided in the UI or via the API. A special API allows users to give feedback for the dialog. Subsequent paragraphs explain individual components in greater detail.

        language sentiment leika synonyms detection handling detection Dictionary Wörterbuch dictionnaire rules sayings wise



      keywords user-expert translations dictionary analyzer popularity intent Query statistics prediction optimization generate answer combine geo provide candidates retrieval coordinates recommendation context service location topics, descriptions user user descriptions session (tree-structured) feedback Fig. 2.

        System architecture

        4.2 Dialog Management The dialog management component handles the user input and manages the user sessions. Thus, it is the central component that predicts the user intention based on the past dialog, the context, and the user input. When a user input is received, the dialog component classifies the input and decides what component is best-suited for handling the user question. The analysis comprises prediction, whether the user has a follow-up question (referring to a previous question) or the user wants to switch to a new topic. Dependent from the specific dialog state, the probabilities of the user intentions (with respect to the user input) are computed. E.g., if the user starts a new dialog, there is a high probability that the conversation starts with greetings or with a general question referring to a service or a location. Follow up questions are typically related to details such as service fees or opening hours of an office.


      A. Lommatzsch

        The intent detection computes a large collection of features. It predicts the input language, checks for location indicators (postal codes, geo-coordinates, district names), frequently used small talk phrases, follow-up questions indicators (related to locations or services) as well as explicit negative sentiments. Based on the intent prediction, the user input is delegated to the best suited components. Finally, the dialog management component combines the information returned by the selected components and generates an answer. In addition, it provides a recommendation for follow up questions or hints how to reformulate questions in order to get optimally matching answers.

        4.3 Small Talk Component The objective of the small talk component is simulating a natural conversion. The small talk component handles greeting phrases and closing phrases as well as question related to appearance, age, name and mood. A small talk dialog does not provide helpful information, but improves the trust in the chatbot. In order to imitate a “natural conversation” this functionality is expected by many users. Looking at log files, we have noticed that many users frequently engage with the small talk component. In addition, the small talk component hints what questions promise valuable information.

        4.4 Questions Related to Recent Events Questions concerning recent events cause remarkable traffic. Users expect short answers or links to relevant documents. Examples for recent events include ser- vice disruptions, technical problems with software, electricity outages, and elec- tions. Our system employs a set of regular expressions to detect this kind of question. In most cases, it merely takes two answers to satisfy these information needs. The set of patterns is continuously updated on new user input.

        4.5 Foreign Language Support Users often expect chatbots to answer questions in different languages. This mandates that systems can detect the language and switch to a different set of answers. This can be done either by a designated agent or a translation service.

        Our system only supports German questions at the moment. This is due to the fact that the underlying knowledge base content is completely written in German content (ensuring the reliability of the answers). If the user input is in another language, the chatbot answers bilingual in German and in the detected language saying that only questions in German can be answered. The language detection is implemented using word and character statistics that are specific for every language. Currently, the system is able to detect German, English, French, and Spanish user input. The component provides an API for integrating an automatic translation service.

        A Next Generation Chatbot-Framework for the Public Administration 135

        4.6 Knowledge Base Mapping Our chatbot has to deal with a large collection of relevant citizen services, loca- tions, and external documents. Thus, it needs an efficient matching of user input to knowledge base entries. This mapping has to support large knowledge bases, handle misspellings, map location information as well as differences in users’ language.

        We have built the core matching component on Apache SolR. First, we have to identify which service, location or topic matches the users’ intents. The system incorporates a multitude of heterogeneous knowledge sources including synonym lists, keywords defined by experts, statistics, and geo-location data to determine potentially relevant concepts and to create optimal queries. Having observed user questions, we update the synonym list to incrementally improve the query creation. In addition, we apply a POS-Tagger to the user questions. The system ranks potentially relevant concepts by popularity and the weights of the query terms. Having determined the most likely service or location, the system selects a sub category. Subcategories include fees, opening hours, and forms. The chatbot replies with detailed information, e.g., describing what fees the user would have to pay. If no subcategory applies, the bot responds with general information about the service. Thereby, the system can adjust the answer to the context. The integration of geo-knowledge bases enables the search for the nearest office or the location in a desired area (specified by district name, postal code, or geo-coordinates).

        This multi-staged processing of user questions, the context-aware transforma- tion and enrichment of questions, and the efficient full-text index-based content selection ensure both a high result quality and the scalability of the system.

        4.7 Recommender In information retrieval and question answering system, one main problem is that users need already domain knowledge for formulating good questions. Thus, an intelligent chatbot should support the user in finding the relevant questions. Therefore, our system comes equipped with a suggestion component. The com- ponent takes the chat history, contextual information, and observations from system usage to suggest questions. Users can “accept” suggestions by click- ing. This minimizes users’ efforts and guides them through complicated service descriptions.

        The component for providing recommendations guides users not familiar with a topic through all relevant aspects of a problem. Users who know what informa- tion is needed can ask directly for the relevant aspect. So, different expectations with respect to the dialog are fulfilled.

        4.8 Data Import Management Our chatbot must aggregate data from heterogeneous sources. Some domains offer structured data, although, formats can vary. Frequently, data quality issues


      A. Lommatzsch

        demand additional processing. For instance, this may require us to validate fea- tures or aggregate data from multiple sources (due to empty or misused fields).

        Our chatbot system provides components for importing JSON, CSV, and XML data. Further, it validates geo-coordinates, addresses, and business hours among others. The chatbot connects these data sources to the internal ontology. The linking capability facilitates customization and extension. Finally, the chatbot comes equipped with a scheduler which periodically updates the ontology.

        4.9 Answer Generation The chatbot should imitate a human agent. Thus, the information must be wrapped into natural language sentences. For this purpose a large set of tem- plates is defined. Typically, several different templates for answering a specific question exist. Randomly selecting a template for generating an answer simu- lates a more human-like behavior ensuring a variance in used sentences. Since the knowledge base (used by the chatbot) has originally been designed for a web page and not for a conversational bot, a tuning of the templates is needed to con- ceal potential mismatches. In addition, the level of providing explanations must be defined in order to ensure compact answers but providing enough details.

        4.10 The User Interface

         . Users encounter a dialog box. Questions and answers are visualized as speech balloons.

        Users enter free text and click on links. Besides, the chatbot includes identifiers to options which users can type in to accelerate communication. The GUI renders html-iframe to simplify integrating the chatbot in existing web portals.

        In addition to this, users can interact with the chatbot via a JSON-based REST interface. Thereby, the chatbot can be seamlessly integrated with other chatbot frameworks. The integration in Alexa is planned.

        4.11 Discussion The developed chatbot architecture has been designed to efficiently handle dif- ferent types of user requests and to deliver detailed answers for a wide spectrum of questions. The system checks whether users continue a dialog about a spe- cific topic or they want to change subjects. Integrating a multitude of knowl- edge sources resolves uncertainties, for instance, by misspellings or using unusual words. The system continuously learns as it observes how users react to answers.

        This extends the initial knowledge base with respect to the user language and the probabilities of requested information pieces. Additional knowledge sources can easily be integrated due to the open architecture. Furthermore the chatbot system can be deployed in new scenarios (e.g. customer service). 4

        A Next Generation Chatbot-Framework for the Public Administration 137 Fig. 3.

        An exemplary dialog.

      5 Evaluation

        The developed chatbot system has been deployed and linked on the service


        portal of the Berlin administratiThis allows us to evaluate the system with real users. Our system went online without announcement in July 2017 as an additional communication channel for citizens. In the first weeks, we observed a constantly increasing interest in the system. In January 2018, the chatbot handled more than 2,500 dialogs. The usage statistic is visualized in Fig.


        system runs without problems and scales well with the number of users ensuring short response times. User Interests.

        We have analyzed which topics users most frequently requested. Figure

         shows the types of questions. The majority of questions relate to ser-

        vices. We expected this as providing answers about services represents the sys- tem’s main objective. Small talk represents a small fraction of questions. Still, users frequently asked these questions. The Requested Services in Detail.

        We have analyzed which services the chatbot most frequently matched to user questions. Table

         shows the statistics. Com-

        paring the statistics with services’ popularity on we observe a high concordance. The chatbot receives slightly more questions concerning “for- eigners”, for instance, applying for citizenship or civil marriage for foreigners. 5


      A. Lommatzsch

        Fig. 4.

        The usage statistic for the chatbot.

        Table 1. The most frequently requested services.

        Rank Service name

        1 Registration of a flat

        2 Apply for ID card

        3 Apply for passport

        4 Marriage for German Citizens

        5 Change flat

        6 Criminal record certificates

        7 Naturalization Fig. 5. The composition of the detected

        8 Marriage user intents.

        9 Registration card

        10 Deregistration of a flat

        The disproportionately high fraction of questions coming from foreigners seems to indicate that foreigners are more familiar with the new “channel” chatbot. Moreover, foreigners may prefer the chatbot over calling the hotline, if they speak German less fluently than necessary. In the analysis of the user questions, we observe a significant fraction of questions asked in English. As the knowledge base contains exclusively explanations in German, questions in other languages are answered by the hint, that the system currently only supports German. In integration of a translation service is planned for the near future.

        Question-Answering Matching Quality.

        We have analyzed how well the chatbot’s answers match users’ questions. We have found that the system has answered ≈85% of questions adequately at the first attempt. ≈9% of the questions referred to topics not scope of the system; for ≈3% of the questions the user intention was unclear for the human annotator that it has been impossible to decide what the user expected (e.g., words or characters without context, such as “street”). For the remaining questions the system provided a valid answer not directly related to the question. A Next Generation Chatbot-Framework for the Public Administration 139

        The three most common reasons why a question cannot be answers immedi- ately are: (1) The wording in the question strongly differs from the “formal” service description. Here the chatbot must learn the “translation” of the user language to the “administrative” language. This can be seen as a type of “cold start problem.” The longer the system runs the more mappings it learns. Based on the season, new concerns/problems may get relevant for certain services (e.g., rats for the service “infection protection”).

        (2) Users ask for meta-information related to the offered services. Typical questions are related to problems in the appointment booking process and the procedure if a transaction key got lost. Running the chatbot we found that the knowledge base provides detailed information about the services; but information how to proceed in case of problems in the appointment booking process and specific exceptions are missing. We extended our knowledge base with frequently requested knowledge to address this issue.

        (3) Several users ask for concrete administrative deeds and provided detailed personal data. As the chatbot is designed for explaining administrative services in general, the chatbot replies to the user in this case, that the bot cannot access their individual reference numbers.

        Analyzing User Feedback.

        In contrast to a search engine, a chatbot system allows users to give feedback about the result quality and the experiences with the ser- vice. In the web application we implemented feedback buttons for this purpose. In contrast to our expectations, the feedback buttons are hardly used—users tend to rather give feedback directly in the chat. Thus, the chatbot must detect, that a user gives feedback to ensure that a suitable response is generated and the feedback is reliably logged.

        The analysis of user feedback helps us to detect shortcoming. We have used negative feedback about incorrectly matched questions and unavailable appoint- ments by retraining the matching policy. Missing free appointment slots are not a problem of the chatbot, but it must be able to explain how to proceed if no appointment slot is available in the desired timeframe.

        Another problem is caused by questions about uncovered domains. As our knowledge base merely provides information about Berlin’s citizen services, ser- vices offered by the Federal Government cannot be answered well. Users often do misunderstand this fact and ask why the responsibilities are so complex. In these cases, the chatbot can merely answer that “why questions” are not the objective of the chatbot.

        Overall, the analysis of the dialogs and the user feedback gives valuable insights on what information should be explained in a better way. The user feedback helps us to detect shortcomings in the service descriptions (e.g., how to change a booked appointment) as well as problems with related systems (e.g., unavailability of appointment slots). Discussion. Overall the developed system shows a good performance provid- ing reliably highly accurate answers related to the wide spectrum of services


      A. Lommatzsch

        offered by the Berlin administration. The analysis of the user input allows us to incrementally improve the mapping from user language to the matching ser- vices. Moreover, the chatbot detects the information important but missing in the knowledge base to overcome shortcomings with existing web portals.

        The chatbot is a powerful additional channel for getting answers related to administrative services. We have found that currently foreigners seem to prefer this channel; but with the increased popularity of chatbots and the high quality of the provided answers, the developed digital assistant will attract more users.

      6 Conclusion and Future Work

        In this work, we presented our chatbot system exemplarily implemented for answering all types of questions related to e-government. The system integrates a collection of heterogeneous data sources, including frequently used general conversion dialogs and service description from several different departments. The system implements a flexible architecture allowing us to integrate additional sources and frequent data updates. In contrast to most other existing chatbot systems, the system provides answers to complex problems—users can ask follow- up questions that are answered with respect to the context and the previous conversation. In our scenario, users get specific information related to costs, opening hours, and the required documents. The system delivers the data in small size “packages” optimized to the users’ needs avoiding an information overload. In addition to answering questions, the system allows users to book timeslots, if they need certain services.

        The system is online on Berlin’s central Service Portal and evaluated with real user feedback. Due to the advantages of the implemented system, the system


        Future Work.

        Currently, we use the feedback collected in the evaluation to dynamically enhance the chatbot. We aim on detecting seasonal trends (e.g., election causing a special interest in election-related services) and improving the mapping from user-defined questions to expert-defined service descriptions.

        In addition to the deployment on Berlin’s service portal, the chatbot will be adapted for additional German regions. This requires the support of region- specific language and regionally relevant concepts.

        Moreover, we work on improving the Named Entity Disambiguation by tak- ing into account the context for ensuring a higher question-answering precision.


      Acknowledgment. We thank the ITDZ Berlin for supporting the development and

      the optimization of the chatbot framework.


        A Next Generation Chatbot-Framework for the Public Administration 141 References


      1. Abdul-Kader, S.A., Woods, D.J.: Survey on chatbot design techniques in speech

      conversation systems. Int. J. Adv. Comput. Sci. Appl. 6(7), 72–80 (2015).

      2. Allam, A.M.N., Haggag, M.H.: The question answering systems: a survey. Int. J.

        Res. Rev. Inf. Sci. 2(3), 211–221 (2012)


      3. Bi, Y., Deng, K., Cheng, J.: A keyword-based method for measuring sentence

      similarity. In: Proceedings of the 2017 ACM on Web Science Conference, WebSci 2017, pp. 379–380. ACM, New York (2017)


      4. Chia, A.: 5 reasons to use the bot, March 2017. Blog of Singapore Govern-



      5. Cui, L., Huang, S., Wei, F., Tan, C., Duan, C., Zhou, M.: Superagent: a cus-

      tomer service chatbot for e-commerce websites. In: Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver, Canada, 30 July–4 August, System Demonstrations, pp. 97–102 (2017)


      6. Deutscher Bundestag. Gesetz zur F¨ orderung der elektronischen Verwaltung sowie

      zur ¨ Anderung weiterer Vorschriften. Bundesgesetzblatt (German Federal Law Gazette), Section 1, 25 July 2013


      7. Kaid, L.L., Holtz-Bacha, C. (eds.): Encyclopedia of Political Communication, Part

      I, chapter E-Government, pp. 200–204. SAGE (2008). ISBN 978-1412917995

      8. Magistrat der Stadt Wien. WienBot - Der Chatbot der Stadt, June 2017. https://


      9. Mishra, A., Jain, S.K.: A survey on question answering systems with classification.

        J. King Saud Univ. Comput. Inf. Sci. 28(3), 345–361 (2016)

      10. Qiu, M., Li, F.-L., Wang, S., Gao, X., Chen, Y., Zhao, W., Chen, H., Huang, J., Chu, W.: AliMe chat: a sequence to sequence and rerank based chatbot engine.

        In: Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pp. 498–503. Association for Computational Linguistics (2017)


      11. Zhai, C., Massung, S.: Text Data Management and Analysis: A Practical Intro-

      duction to Information Retrieval and Text Mining. Association for Computing Machinery and Morgan, Claypool, New York, NY, USA (2016)


      Experimenting a Digital Collaborative

      Platform for Supporting Social Innovation

      in Multiple Settings

        1( )



        Thomas Vilarinho , Ilias O. Pappas , Simone Mora , In`es Dinant ,

        2 B



        3 Jacqueline Floch , Manuel Oliveira , and Letizia Jaccheri 1 SINTEF, Trondheim, Norway


      2 Farapi, San Sebastian, Spain 3

      Norwegian University of Science and Technology, Trondheim, Norway




        Social Innovation is gaining popularity as an approach to

      address societal challenges. Governments, charities, NGOS and organiza-

      tions are taking up the role of facilitating citizens participation into social

      innovation initiatives. Digital collaborative platforms have a great poten-

      tial for enabling and supporting the social innovation process as it facili-

      tates knowledge sharing, cooperative work and networking. In this work,

      we experimented using a digital social innovation platform and associ-

      ated methodology for supporting citizens to do social innovation in three

      different pilots settings: an university course, a contest/hackathon and

      an “in the wild” scenario. We reflect on the participants usage and expe-

      rience with the platform for understanding its added value and uncover-

      ing important considerations for designing and implementing this type

      of platform. The analysis of the experiments highlights (1) the value

      of facilitating collaboration with beneficiaries and across different back-

      grounds, (2) the importance of actively engaging participants on process

      and (3) the needs of adapting the platform for handling complexities

      risen from the social innovation process on real settings.

        Keywords: Social innovation Collaborative platforms ·


      1 Introduction

        Social innovation refers to the development and implementation of innovations (new products, services and/or models) creating value primarily to society, mak- ing social impact and solving a societal challenge. It does so in an inclusive way, having the society, represented by citizens and beneficiaries, playing the role of innovators

        . Social Innovations are not restricted to NGOs and social orga-

        nizations, they can be led by individuals, entrepreneurs, SMEs, governmental

        Experimenting a Digital Collaborative Platform 143

        bodies or any kind of organization willing to make social impact. Existing Social Innovation methodologies and definitions anchor the Social Innovation Process (SIP) into the collaboration and participation of various stakeholders


        Digital platforms are promising tools to support the SIP due to their capabili- ties in facilitating collaborative and crowd-based cooperation, knowledge sharing and networking. Collaborative digital platforms have had success in mobilizing crowds of users for different purposes such as outsourcing (crowdsourcing


        collectively funding entrepreneurial initiatives (crowdfunding

         ]), or sharing

        items and services (collaborative consumption

        ). The establishment of the Col-

        laborative Awareness Platforms for Sustainability and Social Innovation (CAPS)


      ] program in Horizon 2020 highlights the expectations of ICT platforms to

      effectively support social innovation as well.

        Collaborative digital platforms have been used for supporting the innovation process within companies

         ]. How-

        ever, there are very few studies analyzing the effects of using those platforms in the context of social innovation. The studies we found have been rather indirect by relying on the evaluation of experts that did not participated on the SIP facil- itated by the platform

        or solely on data that was recorded in the platform



        Hajiamiri and Korkut

         ] breaks down the value of platforms to support

        innovation through crowd-based design in the following values: supportiveness, collectiveness, appreciativeness, responsiveness, trustworthiness, and tangibility of outcome; where the first three are interrelated. Paulini et al.

         ] identifies,

        among the 3 collaborative ICT platforms they have studied, a few mechanisms to produce successful innovations: (1) supporting communication, in special dur- ing ideation and evaluation, for strengthening ideas, (2) clarifying tasks and roles of those who would assume those tasks and (3) structuring the process. Fuge and Agogino

        uncovers that in OpenIdeo, a digital platform support-

        ing social innovation, the majority of the users would partake into one social innovation initiative facilitated by the platform, but cease to participate after such initiative is finished instead of joining or contributing to other initiatives. They suggest that the active role of community managers to engage participants together with spacing initiatives in time could help to retain users participation. Finally, Ahmed and Fuge

         ] analyzed the winning ideas in OpenIdeo and noticed

        that they were the ones with the highest community engagement and uptake of community feedback into the ideas concepts.

        While the above analysis are valuable, they do not uncover what is the value for the participants in using such platforms for doing social innovation neither how the elements supporting the SIP are used by users in practice and how those affect the SIP. Therefore, in this paper, we focus on users of a digital platform to support social innovation, SOCRATIC ], for analyzing its effect on the SIP. We investigate the following Research Questions (RQs): (RQ1) How do people use the SOCRATIC platform and methodology in practice? (RQ2) What value do people get from SOCRATIC for conducting the SIP?

        144 T. Vilarinho et al.

        The remaining of this paper is structured as follows: Sect.

         introduces the social innovation platform and methodology investigated in this research; Sect.

         presents the research methodology used to investigate the platform; Sect.

         concludes and sug- gests directions of further work based on the results found.

      2 The SOCRATIC Platform and Methodology

        The baseline for performing this research is SOCRATIC, a digital platform and methodology ensemble to support social innovation. SOCRATIC is intended to be applicable to different domains and different types of organizations willing to facilitate the SIP. SOCRATIC was developed based on the needs and practices along the SIP of two distinct organizations, Cibervoluntarios (Cib) and NTNU, for achieving the desired flexibility and applicability


        The SOCRATIC Methodology

         ] presents a SIP heavily inspired by The

        Open Book of Social Innovation

        . The SOCRATIC SIP steps which have

        been covered in the experiments are the following ones:

      • – Preparation: this step marks the set-up of the social innovation environ- ment. It is when Coordinators, representatives of organizations facilitating the SIP, define theirs curated spaces. Coordinators introduce their vision, specific guidelines and how they are able to support the SIP.
      • – Prompts: this step consists in identifying and understanding the societal prob- lem to be solved. The step is lead by a Challenge Owner (CO), an individual or organization deeply interested in solving the societal challenge and who is willing to support innovators solving it.
      • – Ideation: this step is about the generation and development of ideas by Chal- lenge Solvers (CSs) for addressing the societal challenge.
      • – Prototyping: this step concentrates on iteratively materializing the idea, in a lean approach, so that the materialized concepts can be assessed early and refined.

        The SOCRATIC platform supports the methodology by providing a digital meeting place connecting the different stakeholders and facilitating communi- cation and knowledge sharing. The platform features and user interaction flow were designed to guide users through the SIP steps and to include beneficiaries in the underlying activities of each step. Including the beneficiaries in the process is a key aspect of the SIP. The Preparation step has been carried out outside of the platform in our pilots by having the pilot organizations setting up their platform instances and the context for social innovation. The remaining steps (Prompts, Ideation and Prototyping) are directly supported in the platform as described in the next paragraphs.

        In the platform, the SIP is driven by a logic flow where: first, Challenges are defined by a CO describing the societal problem and enabling participants to discuss it, and refine the understanding of the challenge. Then, when the

        Experimenting a Digital Collaborative Platform 145

        Challenge is well-defined, the CO starts the Ideation step. During the Ideation, participants can play a role of Challenge Solver Leaders (CSLs) by creating ideas in the platform that others can contribute to, becoming Challenge Solvers (CSs). The definition of both Challenge and Ideas are done via specific Challenge and Idea templates, while the contributions are done via commenting (as illustrated in Fig.

        . The CO and CSL can edit the challenge and idea, respectively, in order to incorporate the feedback given in the comments.


      Fig. 1. Example of a challenge from the CiB pilot (UI elements highlighted in red)

      (Color figure online)

        At a pre-defined deadline, the CO stops the Ideation and selects the best ideas. They can open up the ideas for voting in order to get feedback from the SOCRATIC community, the platform users, on the feasibility and impact of

        146 T. Vilarinho et al.

        the ideas. The CSLs of the selected ideas can, then, convert them to projects. Projects are represented in the platform through a Project page. The platform automatically suggests the CSL to invite the contributors to join the project and helps finding possible team members within the community, by searching through the community based on skills and interests. In addition, the platform allows the team to plan prototyping iterations and define theirs business model. The prototyping iterations, business models and project description are made available to the community for feedback and discussion.

        Some features of the platform provide orthogonal support to all different SIP steps such as: definition of user profile including skills and interests, recommen- dation of initiatives (challenges, ideas and projects) based on skills and interests matching, possibility to invite internal and external users to join an initiative, messaging, possibility to “like” initiatives and share them on social media, and, finally, it enables COs and CSLs to create “call-to-actions” where they explicitly request for a specific support to their initiative. Along with that, the differ- ent SIP steps supported by the platform are illustrated with guidelines coming from the SOCRATIC Methodology. At last, the platform presents an illustrative video and a link to the SOCRATIC Handbook. The Handbook describes the SOCRATIC SIP through an easy-to-understand approach in order to educate newcomers to contribute to social innovation using SOCRATIC.

      3 Research Methodology

        We have experimented with SOCRATIC in the context of three pilots: (1) the Experts in Teams course, (2) a Social Innovation Contest and (3) an “in the wild scenario”. The different nature of the pilots allowed us to test SOCRATIC along multiple factors such as with/without strong coordination, among younger and older participants and with different extrinsic motivators involved.

        Our research approach was structured on a set of methods belonging to real- world research

        and involved primarily five steps: (i) the elicitation of the

        RQs, (ii) running the pilot scenarios, (iii) collecting data, (iv) analyzing the data for each pilot and (v) aggregating findings from all the pilots and discussing the similarities and differences between the results.

        We used several methods for gathering pilot data: observations during phys- ical gatherings promoted by the pilots, analysis of the data registered by the participants in the platform and semi-structured interviews with participants. Observations have been used in qualitative research for gathering data about participants, theirs interactions, cultures and processes

         ]. All the three pilots

        counted with physical sessions that served to introduce SOCRATIC, trigger collaboration and support the pilots’ participants. During those sessions, the authors of this article or key informants, acted as neutral observers noting aspects relevant to the session such as: the participants understanding of the platform and the SIP, the participants interaction during the sessions and the constraints that could have affected the session. The platform data analysis consisted in going through the public registries of users communication and participation

        Experimenting a Digital Collaborative Platform 147

        during the SIP of the innovations started in the platform. For the semi-structured interviews, we defined an interview protocol aligned with the RQs, described in Sect.

        and including questions about the participants involvement on the dif-

        ferent steps of the SIP, the support given by the platform in those steps and the cooperation patterns between participants.

        Although the data collection strategy and methods were the same in the three pilots, their implementation was adjusted to suit theirs different characteristics and scope. The semi-structured interviews, for example, were adapted to cover the level of usage of the platform and progress achieved in the SIP by the partic- ipants. The observations were tailored accordingly with the goal of the physical session: introduction workshop, facilitation of dialogue with beneficiaries, idea selection, etc.

        The collected data were reviewed using thematic analysis. Thematic analy- sis is a method for analyzing qualitative data according to existing themes and patterns within the data corpus

        . Themes were defined inductively and itera-

        tively influenced by both the RQs and the final structure of the interviews and observations. Overall, this is how the analysis was performed:

        1. Observers transcribed the observation notes from the physical sessions, and interviews were recorded.

        2. Interviews were listened again and coded. The researchers did that by noting data items (interviewee statement or observation) relevant to the research and setting a code to it. A code is a word or short text that express the data feature of interest for the analysis

         ]. At least two different researchers were involved in this step.

        3. Codes were grouped into common themes that explains or formulates evi- dences related to the RQs.

        4. Finally, the datasets were analyzed together within the themes for generating generalized propositions helping to answer the RQs.

      4 The Pilot Scenarios

        In this section we describe each pilot scenario and their key results. Table

         summarizes the pilots.

        4.1 Experts in Team Experts in Teamwork (EiT) is a MSc course taught at NTNU in which stu- dents develop interdisciplinary teamwork skills. Students work in interdisci- plinary teams and establish a project to solve a real-world problem. During the spring of 2016 the EiT course was given having as theme “ICT-enabled Social Innovation for Social Good” and having participants using SOCRATIC to follow the SIP. The course focused on the Ideation and Prototyping steps of the SIP, although the SOCRATIC platform only supported the Ideation at that time.

        The course staff acted as Coordinators and adapted the course structure to use the SOCRATIC platform and methodology as in the Preparation step.

        148 T. Vilarinho et al.


      Table 1. Pilots overview

      EiT SIC Ciberplus Organizer NTNU NTNU Cib Context University Course Contest and “in the wild”


      Coordinating Lectures, facilitating Talks, facilitating platform and SIP

      actions access to COs and access to COs and presentation and

      on-site support support at the triggering the hackathon process COs Autism society Autism society Drawn among workshop participants

        Duration 4 weeks 6 weeks 4 months Nb of challenges, 5, 15, 5 5, 4, 2 87, 9, 0 ideas, prototypes Nb of CSs and 26, 26 12, 10 270, 10 interviewees

        They provided short lectures about the SIP and interdisciplinary work during the course and were available to support the students. The Autism Association of Trondheim (Norway) acted as COs, where 5 people from the association actively collaborated in all the process, inclusively bringing 10 beneficiaries to comment and feedback on the initiatives. The COs described 5 Challenges in the platform after discussing them, as per the Prompts step, with the Coordinators.

        The students, CSs, were involved from the Ideation phase. A total of 26 stu- dents from all different academic backgrounds participated in the course. Despite being held in Norway, the course was in English and 62% of the students were for- eigners. The Coordinators grouped students as to mix nationalities and education background at the beginning of the course. There were a total of 6 groups, each containing 5–6 students. Groups were responsible for describing ideas towards the challenges defined by the COs and for prototyping a selected idea. They had five days to go through the Ideation step and ten for the prototyping.


        The pilot generated a total of 15 ideas and 5 prototypes. There was a lot of interaction between COs and CSs and also among CSs during the process. Not all interactions were recorded in the platform since participants worked both physically together and on-line. Still, all ideas received comments, on average four per idea. Even if there was no incentive for groups to comment and feedback other group’s ideas, they did so. The COs were positively impressed with all the developed prototypes.

        Both students and COs found the process and templates helpful in guiding them and defining the initiatives. They highlighted that the process fostered

        Experimenting a Digital Collaborative Platform 149

        reflection and improvement of the ideas by supporting collaboration. The results of this pilot are further explored in


        4.2 Social Innovation Contest The second pilot happened in a contest setting, the Social Innovation Contest (SIC), in the middle of 2017. The pilot was led by NTNU who was interested in observing how SOCRATIC would perform in a different context and towards a different audience. The Preparation step consisted in developing the concept and timeline of the contest, inviting participants and planning two facilitating events: a on-boarding workshop and a Hackathon. The same NTNU personnel which acted as Coordinators in the EiT pilot coordinated the SIC. They invited experts in social innovation to present talks during the workshop and support the participants during the Hackathon. Six members of the Autism Association of Trondheim played the role of COs defining the challenges and involving bene- ficiaries in the process. CSs were recruited by advertising the SIC in social media and different innovation hubs in city of Trondheim. Finally, twelve participants joined as CSs. They came from different backgrounds and from ages ranging from 20 to 60 years old.

        As in the EiT pilot, COs worked together with beneficiaries and Coordina- tors to define the challenges. They discussed and refined them in the platform before the SIC officially started. The SIC started with the on-boarding workshop where Coordinators, experts and COs presented the SIP, the platform and the challenges. CSs were divided in three groups of four participants and started the Ideation. The Ideation continued after the workshop for two weeks, where participants used the platform to refine ideas. Then, they met again for a two- day Hackathon. The Hackathon started with the selection of the best idea of each group so they would work on it intensively for two days. During those two days, they developed early prototypes, along with business canvas and plans on how to make their solutions sustainable. The best solution was elected by COs and experts and awarded a prize of 10.000NOK (1.200EUR) cash to be used to further developing it into a social startup.


        CSs focused on two of the five challenges and elaborated 4 ideas. The ideas came up during the workshops, but were largely developed via the platform during the time participants were not collocated. The groups used other collab- orative tools (such as google-drive) for collaborating among themselves; and the platform, via editing and commenting the Idea, for collaborating with benefi- ciaries and COs. The ideas got between five and ten comments each. The pro- totypes were built through digital and physical mockups during the Hackathon, supported by a business model canvas. Experts and beneficiaries were available in periods of the Hackathon for giving feedback to the CSs. The final resulting prototypes were rated, by both CSs and COs, as very successful and of high relevance to the beneficiaries.

        Since the COs were the same as in the EiT pilot their experience with the Prompts was straight-forward and similar to the previous pilot. CSs found that

        150 T. Vilarinho et al.

        SOCRATIC helped them going through the SIP and fostered collaboration, in special the close contact with beneficiaries helped improving the Ideation out- comes. Yet, lack of time and features to support groupwork impacted on partic- ipants’ experience.

        4.3 Ciberplus Ciberplus was a pilot led by Cibervoluntarios (Cib). Cib is a spanish non-for- profit organization engaging volunteers on using ICT for social good and social training, courses and online campaigns helping populations with little technol- ogy literacy. Cib would like to use SOCRATIC for supporting theirs volunteers and interested participants for going beyond punctual actions, by doing social innovation. Within its current organizational model, Cib has a severe workload contacting parties and facilitating the actions to happen. They would like, with SOCRATIC, participants to be able to go through the process more indepen- dently, relying less on their role as Coordinators and more on other participants.

        Therefore, they opted to run a more “in the wild”

         ] pilot with limited interven-

        tion from Coordinators and no rules set towards the participants except for the boundaries defined on the platform itself. Moreover, differently from the other pilots, there was no extrinsic reward incentive for participation such as grades at a course or a prize for winning a contest.

        For the Preparation step, Cib adapted the platform and introduced it to the participants through a series of workshops. The platform adaptation consisted in translating it to Spanish and adapting the look-and-feel to match Cibs’ visual image. In addition, the platform was on continuous development during the pilot period, allowing bug fixes and introduction of new features. The pilot started in October 2017 and lasted about four months. Cib carried out 13 workshops spread along those months, reaching out 141 participants.

        The workshops worked as a mean to recruit users to the platform, introduce SOCRATIC and trigger the usage of the platform. The workshops were led and moderated by Cib volunteers which were trained on how to use the platform and which used a common baseline presentation. Cib approached universities, high schools, NGOs and companies for hosting and taking part in the workshops.

        Those in the academia were the most interested in participating, therefore the majority of Cib workshops were held in universities and high schools. In addition, one workshop was held together with NGOs representatives.

        During the workshops, the moderator asked the audience if they had a soci- etal challenge they were keen to work on collectively using SOCRATIC. Partici- pants who had a challenge in mind shared it and the other participants decided which challenge to join. The group of participants working on the challenge used the remaining time of the workshop to describe the challenge in the platform and, ideally, would keep using the platform later on. In that sense, any workshop attendant could become a CO, CSL and CS.

        Experimenting a Digital Collaborative Platform 151

        Results. Participants of the Ciberplus pilot mainly used the platform during the workshops. Half of the participants we interviewed said that they did not use it further because the platform was not mature enough, while the others had each one different reasons such as lack of time, perceived lack of a critical mass of users in the platform for obtaining expected support, etc. As a result, the pilot produced many challenges, 87, but very little participants interaction. Challenges received none or up to three comments and the COs did not interact with those who commented. Consequently, most of the initiatives stopped in the challenge definition. Seven cases went as far as the Ideation step: in two cases both challenge and idea were defined during a workshop and in the other cases the ideas were provided by participants out of the workshops. At the end of the pilot, the platform counted with 270 users registered in contrast to the 141 directly reached in the workshops.

        Many of the challenges definition provided by the Ciberplus pilot partic- ipants were actually ideas aiming to tackle a societal challenge. Participants described their idea using the Elevator Pitch section of the Societal Challenge template, instead of describing the challenge and waiting until the Ideation phase to describe theirs idea. It is a bit unclear whether participants did not under- stand the concept behind the Prompts phase or if they wanted to shortcut the process and start from their idea. However, through the interviews, we learned that some of those participants had come to the workshop with existing projects or very well elaborated ideas in mind.

      5 Discussion

        In this section, we present the findings of the thematic analysis of the three pilots under five main themes. The themes’ links to the RQs are illustrated on Table



      Table 2. Mapping between thematic analysis themes and the RQs

      Theme RQ1 RQ2 RQ3 The overall value of digitally supporting the SIP

        X X Value of specific platform components towards the SIP

        X X Process and flexibility

        X X The importance of facilitators

        X X Physical presence

        X X

        5.1 The Overall Value of Digitally Supporting the SIP In the EiT and SIC pilots, CSs collaborated directly with beneficiaries and COs along the process and, consequently, explicitly valued the beneficiaries par- ticipation : “It was really helpful to have someone [referring to a beneficiary] there so we can ask him how he felt [about the ideas].” As another participant

        152 T. Vilarinho et al.

        points, hearing the challenge from a beneficiary was more meaningful then read- ing it from an unpersonalized source: “There’s one thing reading about autism in school papers but actually hearing about it from people who meet the challenges every day was really useful.” The participants also experienced that the bene- ficiaries were very interested on their initiatives and eager to help: “They [the beneficiaries] are really good at responding when we make contact with them.”, “We got our feedback from at least three persons [challenge owners]. They were constructive critiques or constructive thoughts, so they were helpful.”

        As participants from the Ciberplus pilot did not go far along the SIP, we investigated the value of the platform by asking them about the potential value of the platform and the value they got from SOCRATIC during the workshops. The interviewees understood the main value of the platform as the crowd-based collective intelligence enabling gathering feedback and support from others with different ideas and skills. More specifically, they had interest on the following possibilities: (1) finding collaborators to supply the human resource need of a project; (2) finding expertise able to handle specific tasks; (3) getting feedback from those of different backgrounds for improving and further building the ini- tiative; (4) raising awareness about the initiative and (5) measuring the commu- nity support and appreciation of the initiative. Participants leading initiatives valued the feedback from others with different perspectives during the workshops as one of them highlights: “they helped me express it (the challenge) the best way possible, so that people like them, which were not familiar with the goal, could understand it quickly”.

        The value expected by the Ciberplus participants and experienced in the workshops was confirmed in the other pilots where participants had further interaction with the platform. Interviewees from those pilots thought that the collaboration with participants from different backgrounds was very positive, as one points out: “The fact that it is supposed to be an event for people from different backgrounds is really good. I think it was quite effective and I like that people were willing to contribute with their own skills and in a very iter- ative process.” Besides direct collaboration, ideas shared and described in the platform helped inspiring participants to come up and assess ideas as suggested by those two participants statements: “It was easy to find inspiration in other ideas, you might combine some ideas and make a completely new edit!” and “to see that some had the same idea that we had, made us reassured that we were on the right track”.

        The values of supportiveness, collectiveness, responsiveness and apprecia- tivenes identified in

         ] are confirmed during our experiments, though it is

        worth mentioning that participants specially valued that the contributions were coming from real beneficiaries and people with different backgrounds.

        5.2 Value of Specific Platform Components Towards the SIP Participants from the three pilots found the templates useful for describing chal- lenges and ideas. Some participants highlighted that they triggered reflection: “I think to force myself to look at the challenge from different point of views.”,

        Experimenting a Digital Collaborative Platform 153

        while others thought that they helped more clearly describing the initiative. A participant claimed that by using the description of her challenge as in the template for explaining it to her parents they understood it immediately, while, previously, she has not been able to explain them. On the other hand, some participants thought that enabling the inclusion of other description elements such as videos would help even further describing the initiatives: “I think the template is good [. . . ] I just would have liked it if it was another way than just text, [. . . ] if I could have visuals or maybe even video, I think presenting the idea and getting others to understand the idea could be easier.”

        The template fields offered limited space for describing each aspect of the challenge or idea. That design choice triggered mixed feelings between a few par- ticipants. While a participant said it was too restrictive and did not correspond to the level of description he wanted to provide, another thought that it was ideal to describe the most essential aspects of the idea. He saw the description to be provided within the templates as a trigger to gather interest. Therefore, he considered that it needed to be short as people have limited time to read. Just after others are involved, he would then feel that it is important to describe it deeper: “The idea of [the template] being short is good, so that people under- stand it easily, by reading it quickly. After there is interest, then, it would be worth detailing more.”

        Two interviewees highlighted the platform feature of recommending initia- tives based on the matching of skills and interests. One of them suggested it being further integrated by allowing CSLs explicitly describing expertise, skills and resources needed by the initiative and keep track of them through the plat- form. Participants suggested integrating further real-time communication into the platform: “if people would be online at the same time, it would be good to have a chat possibility, where if you see somebody’s already working in an idea, you can establish a messaging communication. That would be very interesting.”

        Some of the interviewees mentioned that it would be interesting to provide additional numbers related to the initiatives in the platform, such as: how many people have read the challenge/idea/project, the percentage of viewers per user profile and the level of activity in an initiative. Those were suggested both as to enable initiative leaders to identify where to further promote the initiative and to provide additional metrics to motivate participants. Related to motivation, the participants of the EiT pilot saw the voting and selection of ideas at the end of the Ideation step as a competition which raised their motivation: “I think it was a competition. I don’t know if the other groups took it as a competition. We were really triggered by it. We immediately wanted to win as a team.”

        The features highlighted by the participants are in line with what was identi- fied as the overall value of the platform. They relate to creating awareness about the challenge and the innovation, helping reach out participants of different pro- files and facilitating the collaboration between users.

        154 T. Vilarinho et al.

        5.3 Process and Flexibility The SIP as implemented in the platform brought some constraints that were not necessarily part of the process. For example, some participants of the Ciberplus pilot came to the workshop with an existing social innovation idea or project, and by using the platform, they were confronted with the need to describe the societal challenge first. In these cases, participants described their idea or project using the challenge template instead. During the interviews, we learned that those were people with a strong drive to take action and who felt to some extent constrained by a well-structured process. Although it is important to have CSs reflecting on the challenge before trying to solve it, the Ciberplus pilot showed that there will be moments where users may start with an existing idea or project. It is important to find ways to enable them starting from a later SIP step, but, at the same time, to ensure that they have covered important aspects of the previous steps (such as identifying root causes of the challenge in the Prompts or verifying an Idea feasibility and relevance in the Ideation).

        Another constraint introduced by the platform was that the leading of initia- tives, and consequently editing rights, was personal and non-transferable. Such constraint finally hindered the co-editing of challenges and ideas description forc- ing the participants to use collaborative tools such as google docs in addition to the platform. It also did not represent the reality, as most of the initiatives that came out of the pilots were led by more then one person and most of the groups of EiT and the SIC adopted a flat hierarchy as one of them states: “We have a quite flat structure. We don’t have any leaders.”

        Still, related to ownership, there were cases where those describing a chal- lenge or idea did not want to take a leading role, but rather thought of his elicitation as contribution itself. They thought that the elicitation could inspire others or eventually be embraced by people willing to take ownership of it. Such role of “seeding” an inspiration was not foreseen and implemented in the plat- form, as passion and drive are crucial for bringing social innovations forward. However, providing some support for the “seeding”concept could foster a wider participation. It can, as well, serve as a basis for a supplementary Ideation for- mat, resembling brainstorming, where wild ideas, that may not be feasible, are encouraged to serve as inspiration to others.

        5.4 The Importance of Facilitators One of the biggest differences between the pilots led by NTNU and the one led by Cib was the level of active engagement of Coordinators in steering and supporting the process. In the NTNU pilots, Coordinators worked with COs to insert the SIP into a “specific context” (the course or SIC) with a defined timing for each of the SIP steps and activities to support the steps. Those activities were done in conjunction with the SIP support provided in the platform, as to strengthen it. For example: NTNU would bring COs and CSs together physically for presenting the challenge posted in the platform, which would increase the empathy between COs and CSs and introduce themselves personally. After that,

        Experimenting a Digital Collaborative Platform 155

        CSs were comfortable to contact COs and further discuss the challenge and ideas. Another example was when Coordinators taught and exemplified the usage of offline tools described in the SOCRATIC methodology, such as brainstorming for Ideation, and, as a result, CSs used the tools and uploaded theirs results into the platform. Meanwhile, in Ciberplus, Coordinators simply organized workshops where they presented the platform to interested groups and expected COs and CSs to emerge, self-organize and cooperate autonomously. Although the platform offers features for self-organization, Ciberplus users did not take advantage of those features. COs did not see or answer to comments to theirs challenges and participants did not come forward to invite others to contribute to the process.

        The results were that despite having more users joining the platform and creating initiatives in Ciberplus pilot, the level of development of the Social innovation initiatives and collaboration during both EiT and SIC pilots was much superior. That influenced the understanding of the platform, the SIP and how much guidance the participants experienced. During the NTNU pilots, the process happened as expected and participants noted the synergy of the platform with the process as two of the participants points: “I think that when you use the platform, you are defining the way you are going to do the process. . . ” and “the platforms really help so instead of just having ideas here and there and you compare them”. Meanwhile, participants from Ciberplus used the Prompts step to describe ideas or projects rather then reflect on the challenge. Moreover, some Ciberplus participants ended up not understanding the platform and the SIP flow on it as one of the interviewees mentioned, and as an observer from one of the workshops noted: “(the understanding) varied according to the user profile.”

        One of the Ciberplus interviewees pointed that he missed having example projects that illustrated the SIP, which showed how participants effectively used the platform and the tools to support the SIP. In fact many of the Ciberplus COs confirmed looking at other challenges as examples before writing their own.

        Those results sustain the conclusions from

         ] regarding the importance on

        the active role of facilitation in order to successfully foster the SIP. That seems specially crucial at the beginning when users are still learning about the pro- cess, platform and social innovation. In long term, it may perhaps be possible to achieve a more self-managed community by having the coordination roles natu- rally spread among community members and illustrated by a significant amount of success stories in the platform.

        5.5 Physical Presence All the pilots included physical meetings. In Ciberplus it happened as an on- boarding workshop, while, at the EiT and SIC, participants worked together collocated during different periods. During the interviews when physical pres- ence was discussed, participants thought that it was essential to meet at least a few times in order to properly cooperate and work together. One of the SIC participants, for example, believes that they would not have been able to come up with the same quality of a result without meeting physically: “I think the

        156 T. Vilarinho et al.

        platform helps but I don’t think it’s a good fix. If we didn’t have the meetings before, in the beginning or in the end, I don’t think we could have come up with as good of idea. I don’t think you can replace the physicality. . . ”. While a Ciber- plus participant pointed that further collaboration requires mutual trust and that trust is more easily established when people know each other physically.

        Indeed collocation of CSs may be crucial for many of the initiatives. Includ- ing a location field in the users profile in the platform could facilitate users living next to each other to self-organize, as an interviewee noticed. Besides that, Coordinators should facilitate physical events as they did in the pilot and engage initiative leaders to do the same.

      6 Conclusion and Future Work

        The pilots show clearly the value of such digital social innovation platform in gathering feedback from different types of users and strengthening the innova- tions. However for establishing further participants cooperation and enabling them to effectively work together, Coordinators need to actively support and engage them in the process. Besides that, co-location seems to be an important determinant of success. At last, this study identified prospect design variations of the platform and pilots: supporting group ownership and enabling innova- tion “seeding”. Directions of future work include testing those variations besides running longer pilots and pilots in different contexts.


      Acknowledgments. The work behind this study was possible thanks to the support

      from the EU-funded H2020 SOCRATIC project OPTET under grant agreement ref.

      688228, the Marie Sklodowska-Curie grant agreement No 751550 and the “Sharing

      Neighborhoods” project funded by the Research Council of Norway under the program

      BYFORSK (contract 270737). Furthermore, we thank the participants of the pilots, and

      the project partners for the collaboration during participant recruitment and discussion

      of the results.


        1. MakeSense. Accessed 6 Oct 2017

        2. OpenIDEO. Accessed 6 Oct 2017


      3. Ahmed, F., Fuge, M.: Capturing winning ideas in online design communities. In:

      Proceedings of the 2017 ACM Conference on Computer Supported Cooperative Work and Social Computing, pp. 1675–1687. ACM (2017)


      4. Botsman, R., Rogers, R.: What’s mine is yours: the rise of collaborative consump-

      tion (2010)


      5. Boudreau, K.J., Lakhani, K.R.: Using the crowd as an innovation partner. Harvard

      Bus. Rev. 91(4), 60–9 (2013)


      6. Braun, V., Clarke, V.: Using thematic analysis in psychology. Qual. Res. Psychol.

        3 (2), 77–101 (2006)


      7. Chamberlain, A., Crabtree, A., Rodden, T., Jones, M., Rogers, Y.: Research in

      the wild: understanding ‘in the wild’ approaches to design and development. In: Proceedings of the Designing Interactive Systems Conference, pp. 795–796. ACM (2012)

        Experimenting a Digital Collaborative Platform 157


      8. Dinant, I., Floch, J., Vilarinho, T., Oliveira, M.: Designing a digital social innova-

      tion platform: from case studies to concepts. In: Kompatsiaris, I., Cave, J., Satsiou,

      A., Carle, G., Passani, A., Kontopoulos, E., Diplaris, S., McMillan, D. (eds.) INSCI 2017. LNCS, vol. 10673, pp. 101–118. Springer, Cham (2017).



      9. Fuge, M., Agogino, A.: How online design communities evolve over time: the birth

      and growth of OpenIDEO. In: ASME 2014 International Design Engineering Tech- nical Conferences and Computers and Information in Engineering Conference, p. V007T07A038. American Society of Mechanical Engineers (2014)


      10. Gordon, P., Fuge, M., Agogino, A.: Examining design for development online: an

      HCD analysis of OpenIDEO using HCD/UCD metrics. In: ASME 2014 Inter- national Mechanical Engineering Congress and Exposition, p. V011T14A017. American Society of Mechanical Engineers (2014)


      11. Hajiamiri, M., Korkut, F.: Perceived values of web-based collective design plat-

      forms from the perspective of industrial designers in reference to Quirky and OpenIDEO. A — Z ITU J. Fac. Archit. 12(1), 147–159 (2015)

        12. Howe, J.: The rise of crowdsourcing. Wired Mag. 14(6), 1–4 (2006)


      13. Kawulich, B.B.: Participant observation as a data collection method. In: Forum

      Qualitative Sozialforschung/Forum: Qualitative Social Research, vol. 6 (2005)


      14. Martinelli, F.: Social innovation or social exclusion? Innovating social services

      in the context of a retrenching welfare state. In: Franz, H.W., Hochgerner, J., Howaldt, J. (eds.) Challenge Social Innovation, pp. 169–180. Springer, Heidelberg (2012).


      15. Mollick, E.: The dynamics of crowdfunding: an exploratory study. J. Bus. Ventur.

        29 (1), 1–16 (2014)


      16. Mulgan, G.: The process of social innovation. Innov. Technol. Gov. Globalization

        1 (2), 145–162 (2006)


      17. Murray, R., Caulier-Grice, J., Mulgan, G.: The Open Book of Social Innovation.

        National Endowment for Science, Technology and the Art London (2010)


      18. Pappas, I.O., Mora, S., Jaccheri, L., Mikalef, P.: Empowering social innova-

      tors through collaborative and experiential learning (2018). To be Presented at EDUCON 2018 - IEEE Global Engineering Education Conference, Tenerife, Spain


      19. Paulini, M., Murty, P., Maher, M.L.: Design processes in collective innovation

      communities: a study of communication. CoDesign 9(2), 90–112 (2013)


      20. Regional, DG and Policy, Urban: Guide to social innovation. European Commis-

      sion, Brussels (2013)

        21. Robson, C., McCartan, K.: Real World Research. Wiley, New York (2016)


      22. Romero, I., Rueda, Y., Fumero, A., Vilarinho, T., Floch, J., Oliveira, M., Dinant,

      I., SOCRATIC Consortium, : SOCRATIC, the place where social innovation ‘Hap- pens’. In: Bagnoli, F., Satsiou, A., Stavrakakis, I., Nesi, P., Pacini, G., Welp, Y., Tiropanis, T., DiFranzo, D. (eds.) INSCI 2016. LNCS, vol. 9934, pp. 89–96. Springer, Cham (2016).


      23. Sawhney, M., Verona, G., Prandelli, E.: Collaborating to create: the internet as a

      platform for customer engagement in product innovation. J. Interact. Mark. 19(4), 4–17 (2005)

      24. Sestini, F.: Collective awareness platforms: engines for sustainability and ethics.

        IEEE Technol. Soc. Mag. 31(4), 54–62 (2012)


      25. Vilarinho, T., Floch, J., Oliveira, M., Dinant, I., Pappas, I.O., Mora, S.: Developing

      a social innovation methodology in the Web 2.0 era. In: Diplaris, S., Satsiou, A., Følstad, A., Vafopoulos, M., Vilarinho, T. (eds.) INSCI 2017. LNCS, vol. 10750, pp.

      168–183. Springer, Cham (2018).

        Innovations and Digital Transformation


      Innovation Management Methods

      in the Aviation Industry

        Karl-Heinz Lüke 1(&) , Johannes Walther 2 , and Daniel Wäldchen 2 1 Ostfalia University of Applied Siences, Wolfsburg, Germany


      2 IPM AG, Hannover, Germany


      Abstract. Against the background of significant changes such as increased air

      traffic, fierce international competition as well as the radically changing political,

      technological and environmental conditions, the aviation industry keeps looking

      for product, process and business model innovations. As a result, the importance

      of efficient innovation management is growing rapidly. This paper discusses

      specific innovation methods which can be assigned to three cooperation fields:

      supplier, customer and in-house. The research project ascertains their particular

      implementation level in the aviation industry. The results are presented in this

      paper. One of the findings is the critical importance of Big Data as an explicit

      in-house innovation method. Its growing significance is a result of the enormous

      data volume increase, which is an exceptional challenge for the industry,

      nowadays and in the future. Consequently, a concept for a reference architecture

      is presented in this paper which will enable the aviation industry to cope with

      the challenges.


      Keywords: Innovation management Open innovation Aviation industry

      Big Data Requirements Architecture concept

      1 Current Situation and Motivation

        Significant changes such as increased air traffic, seat mile cost pressure, new tech- nologies, customization and fierce international competition have been affecting the aviation industry for decades [

        In addition, political and environmental requirements

        such as optimizing fuel efficiency and minimizing emissions keep a constraining influence on aircraft manufacturers (OEMs) and their supply chain [

        The industry

        therefore keeps looking for innovative ways to meet the political, technological and ecological challenges to remain on the winning track and meet the growing demand for new commercial aircraft


        There is a common understanding that innovation is a fundamental requirement to cope with the challenges of an increasingly connected world. It is seen as a key source and force of competitive advantage in a society which is constantly changing [


        do we define innovation? Although no commonly accepted definition of innovation exists, it is safe to say that “Innovation is: production or adoption, assimilation, and exploitation of a value-added novelty in economic and social spheres; renewal and

        162 K.-H. Lüke et al.

        enlargement of products, services, and markets; development of new methods of production; and establishment of new management systems. It is both a process and an


        outcome“ [ While the industry is driven by the constant demand for new ideas, it is investing in research and development. But before new ideas can make a difference, there is a long, complex route and no certainty whether the necessary investments and changes will provide commercial success. Therefore, many companies are making creative changes to established products and services by employing new technology

        To address the

        goal of employing the right innovation with effective measures at the right time or defined innovation process, where internal and external (new) knowledge are linked to the already existing know-how. Helpful innovation management concepts are needed that cover product, process and business model innovations.

        The aviation industry is dealing with highly complex projects with a lot of money, manpower und resources involved

         ]. These large projects are also called programs

        which are involving a large number of related projects [

        Innovation on the other hand

        has a complex nature, too. The task for the industry stakeholders is to address both challenges with strategic tools to ensure successfully reaching the set goals without exceeding the time and financial limits. Adding to the difficulty is the urge to recognize and eliminate complications at an early stage of the innovation process.

        Long and complex supply chains with a few Original Equipment Manufacturers (OEMs) characterize the aviation sector

         ]. Global companies such as Airbus, Boeing

        and Bombardier depend on a large number of suppliers to deliver components and services that meet the required quality standards in a timely, cost effective and well-coordinated manner

         ]. 100% on time and on quality is requested all along the

        value chain and extends to the Aircraft Lyfe Cycle and the Aftermarket. Employing new technology including composites materials and electronic controls needs to be agreed on and tuned along the long and complex supply chain. To benefit from new developments in aviation technology and speed up design and development, the challenge of adopting new organizational paradigms and efficient innovation man- agement tools along the value chain need to be integrative addressed.

        The aviation market report record sales with immensely growing net income [ The Airbus results from 2017, for example, overachieved their goals on all key per- formance indicators [

        The net income at Airbus, for example, increased from € 995

        million in 2016 to € 2,873 million in 2017. With 2,807 million, about the same amount was spent on own research and development (R&D) at Airbus, in 2017 [


        As the numbers proof, the industry is already drawing on a fundamental playbook of continuous improvement. This accelerates change and requires an increasing com- mand of change management

         ]. The sector is keen to view expanding international

        partnerships in the age of globalization, Industry 4.0 or digitization and Additive Manufacturing, for example, as levers to ensure future successful performance. Inno- vation centers are driving the ever so abiding passion of the industry for innovation and technical excellence. Innovation management, therefore, plays a key role in achieving the shared goal of “flying high”.

        Innovation Management Methods in the Aviation Industry 163

        The idea of this study is to take a closer look at the innovation management tools that are already applied in the aviation industry to drive transformation and changes and those which can be transferred from other industries or academia. Changes and inno- vations continue to evolve, that is a fact. How to control, streamline and guide them is the challenge.

      2 Market Insights and Research Needs

        The relevant literature on innovation management is vast

         ]. Researchers in many fields of study have examined innovation as the key means of adapting to change.

        There has been an enormous amount of studies focusing on innovation from all kinds of perspectives such as economists, organizational sociologists, technology and theorists.

        Cossan and Apydin

        for example, consolidated the state of academic research on

        innovation based on a systematic review of literature published since 1983, limiting their sources to peer-reviewed journals. They come to the conclusion that “innovation research is fragmented, poorly grounded theoretically, and not fully tested in all areas”.

        Researchers may have looked at innovation within an organization, as a product or process, as technical or administrative. One of the early studies conducted by Marquis ] focused on the difference between incremental, radical and system innovation.

      Other explore in the direction of product development, open innovation, portfolio management or innovation in certain industries such as automotive

         ]. There are also a vast number of marketing driven contributions in peer-reviewed journals.

        The discourse on innovation found special attention in the automotive industry. There are much more contributions on innovation in the automotive industry from academia and experts than on innovation in any other industry [


        To focus on a whole industry or even sector is a very complex effort which relates to large complex systems. The projects in the aviation industry are more complex than in the automotive industry involving an even longer supply chain. Therefore, complex projects as they are in the aviation industry have received limited research attention.

        No generally accepted framework has emerged so far to support the analysis of highly complex and innovative projects

         ]. Existing research offers limited prescrip- tive ideas on managing innovation.

        To find a comprehensive approach to research innovation in the aviation sector and the importance of the supply chain as a source of innovation is a question worthwhile discussing.

        The opportunity for the aerospace industry is to involve the research community for careful early analysis of innovation tools applied. To take an innovation management perspective and learn from mistakes and set backs will help to prevent, avoid and repeat failures.

        164 K.-H. Lüke et al.

      3 Innovation Methods Evaluation

        3.1 Innovation Management Methods The aviation industry is searching intensively for product, process and business model innovations. Therefore, the meaning of an efficient innovation management is growing rapidly. In this paper, specific innovation methods are discussed which can be assigned ). to the cooperation fields: supplier, customer and in-house (Fig.

        Supplier Customer In-house Co-operation Megatrend Study Employee with Universities Day-in-Life Suggestion External Contest Visits System

      for Ideas Test Market Internal Ideation

      Start-up Field Tests Contest

      Identification and Ideation Internal Ideation

      Co-operation Workshop with Workshop Innovation Lead Users Scenario Scouting Insight Clinique Engineering Value-added Roadmap Networks Development Innovation Big Data Communities

        Fig. 1. Innovation methods

        Supplier Related Innovation Methods Co-operation with Universities: research and development co-operations with com- panies and universities for knowledge transfer, and to initiate a collaborative learning process for product and process innovations. External Contest for Ideas

         ]: the company is inviting e.g. suppliers, universities and research facilities to submit ideas for solving a specific task within a certain timeframe.

        A board is evaluating the ideas based on certain criteria and award them for their performance. Start-up Identification and Co-operation: systematic identification of start-ups (e.g. start-up scouting, start-up pitching, innovation competition) and co-operation on var- ious intensity levels (e.g. general information exchange, supplier relationship, customer relationship, research & development co-operations, API co-operations, corporate investments, Acqui-hires, M & A). Innovation Scouting [

        systematic and continuous search for innovations and

        innovative solutions outside the company to ensure a faster, more creative and focused

        Innovation Management Methods in the Aviation Industry 165

        innovation process. Innovation scouts screen predefined innovation fields and con- tribute ideas to the company.

      • added Networks (Strategic Alliances) [

        value-added networks consisting of

        Value legally independent but economically dependent companies which are linked by a relatively stable connection that allow the company to align their processes more efficiently across different value-added stages.

        Innovation Communities

         ]: a community of like-minded actors, often consisting of

        several companies and institutions that are related to a task and are pushing forward certain innovation projects (e.g. open source).

        Customer Related Innovation Methods Megatrend Study: a megatrend study can be carried out on an annual or biannual basis to identify the impact of sociocultural, political and technological trends in the long run. The study can be used for innovation development.

        Day -in-Life Visits

         ]: observation of customer behavior revealing their demands

        which traditional market research is not investigating. Special functional teams, for example, personally visit the customer and observe them in their environment. The documented results can be used for innovation development.

        Test Market

         ]: test markets can be simulated in virtual environments or in labora-

        tories. The participants experience using products and services. On this basis, a prognosis for purchase intensions can be derived.

        Field Test: innovations are introduced to a chosen local test market of selected test users (e.g., Hamburg, Munich). Ideation Workshop with Lead Users

         ]: ideation workshops intend to delve deeply into the creative potential of participants to develop an ideal product design or service.

        During the workshop, creative techniques such as learning from other brands or building the future with Lego bricks can be used. Participants of ideational workshops can be lead users or a combination of experts and lead users.

        Insight Clinique

        the insight clinique is used for direct interaction with the

        customer and can approach the users with prototypes or concepts of products or ser- vices. The aim is to identify use barriers or latent needs.

        In -house Related Innovation Methods Employee Suggestion System [


        a defined internal business process for the evalua- tion of process improvements and product innovations. Staff gets financial incentives in case their suggestions have an impact on cost savings or product improvements.

        Internal Ideation Contest [


        companies searching for innovations define a specific innovation field which is communicated, for example, online (Intranet or Internet). Staff is invited to provide ideas and suggestions specifically regarding this innovation field. The company awards the winning innovation and decides how to use it in the future.

        Internal Ideation Workshop

         ]: an internal ideation workshop can provide ideas for a

        new product or service. Different creative techniques can be used such as building the future with Lego bricks. Participants of ideation workshops can be experts from dif- ferent divisions and hierarchy levels of the company.

        166 K.-H. Lüke et al.

        Scenario Engineering

         ]: the scenario technique is used to select strategies and to

        raise awareness about probable future events, including soft facts concerning the description of the future. The research framework and the key actors and trends are defined, important uncertain factors are specified, and scenarios are designed, tested and revised. Roadmap Development [

        roadmaps are used in project management, especially for

        projects that take a slightly longer time horizon. Technology roadmaps include the future development of a specific technology or a whole technological industry. Big Data [

        the analysis of large amounts of unstructured or semi-structured data

        (Big Data) is helpful to discover unfamiliar patterns, unknown correlations, and other useful information.

        3.2 Evaluation Criteria The objective of this research project “Innovation Management in the Aviation Industry” is to systematically record the innovation management methods used in the aviation industry and to map them in an innovation radar. For this purpose, a ques- tionnaire was developed in which the participants could indicate the application level of Open Innovation methods described above on a five-level ordinal scale (very high (5) to very low (1)). The results were separately mapped in innovation radars for the cooperation fields: supplier, customer and in-house.

        The survey (online questionnaire in English) was conducted in November 2017. 167 companies from the aviation industry took part (respondents rate 19,6%), most of which operate as OEM, 1st-tier or 2nd-tier suppliers (70%) and generate sales of less than EUR 500 million (75%). The participating companies primarily implement pro- cess innovations (70%), secondary apply product innovations (50%) and finally use business model innovation (35%). More than half of the respondents classified the maturity rate of their innovation management as high or very high.

        3.3 Analysis of Applied Methods/Innovation Radar The results of the research project are mapped in innovation radars [

        for the cooperation fields supplier, customer and in-house.

        With regard to supplier integration (Fig.

        the aerospace industry applies the

        methods “Value-added networks (strategic alliances)” (3,45) and “Co-operation with Universities” (3,41) most frequently. “External Contest for Ideas” (2,54) and “Start-up Identification and Co-operation” (2,71) are of minor importance.

        On the customer side (Fig.


        ), the methods “Day in Life Visits” (3,20) and “Ideation Workshop with Lead Users” (3,09) are paramount, while “Megatrend Study” (2,51), “Test Market” (2,52) and “Insight Clinique” (2,63) are applied on a small scale.

        In-house, (Fig.


        ) the innovation sources “Roadmap Development” (3,70) and “Employee Suggestion System” (3,42) are predominant. Less relevant in the aviation industry are the sources “Internal Ideation Contest” (2,70), “Scenario Engineer- ing”(2,75) and “Big Data” (2,77).

        Innovation Management Methods in the Aviation Industry 167

      Fig. 2. Innovation radar: supplier


      Fig. 3. Innovation radar: customer

        3.4 Customer Segment Evaluation Interesting conclusions can be made by deeper analysis of the data results, featuring the customer segment classes revenue and value chain. The detailed examination of the revenue class (less than 50 million, 50–500 million and more than 500 million) covers the economic aspect, the analysis of the value chain class (OEM and Airline, Tier 1,

        168 K.-H. Lüke et al.


      Fig. 4. Innovation radar: in-house

      Table 1. Distance analysis customer segment “Revenue”: Minkowski metric

      Less than 50 million 50–500 million More than 500 million Less than 50 million 0,00 0,28 0,52

        0,00 0,24 50–500 million More than 500 million 0,00

      Table 2. Distance analysis customer segment “Value Chain”: Minkowski metric

        OEM and Airline Tier 1 Tier 2 Tier 3-n OEM and Airline 0,00 0,41 0,11 0,03 Tier 1 0,00 0,52 0,38 Tier 2 0,00 0,14 Tier 3-n 0,00

        Tier 2, Tier 3-n) focuses on the position in the supply chain. The following calculations (see Table

         for the customer

        segment value chain class) can be made for the distances between each customer segment in terms of revenue and value chain class, using the well-known Minkowski metric. The Minkowski metric is a popular method to calculate the distance mea- surements when a metric structure of variables is given. This method is applied e.g. in the cluster analysis [

        Innovation Management Methods in the Aviation Industry 169 " # X J r r 1

        d x x k ; l ¼ kj lj ð1Þ j


        d k ; l : distance of the objects (customer segments) x ; x : value of variable j with objects k ; l j ð ¼ 1 ; Þ, here: J ¼ 1 kj lj . . .; J Minkowski constant, here: r r 1: ¼ 2

        The mean value of each customer segment (e.g. OEM and Airline, Tier 1) within a of Table

         allows the following observations:

      • The customer segments of the revenue class less than 50 million and more than 500 have the highest distance measurement (0,52). This implies that the seg- million ments differ referring to the mean value when applying the different innovation methods determined in this survey.
      • The result of the customer segment comparison 50–500 million and more than 500 million shows the lowest distance measurement (0,24) in this survey. The distance measurement is significantly lower than in the other segments.

        The following observations can be made regarding the customer segment value chain class Table


      • The customer segments Tier 1 and Tier 2 have the highest distance measurement (0,52). They differ referring to the mean value, significantly.
      • Obtaining a high distance measurement (0,41), the customer segments OEM and Airlines and Tier differ referring to the level of application.
      • The lowest distance measurement (0,03) can be observed between the customer segments OEM, and Airlines and Tier 3-n. This implies that the distance measurement of these customer segments is much lower than within the other customer segments.


        and (b) gives a detailed comparison of the highest distance measurements

        in each customer segment class. Within the segment class “revenue”, it can be observed that the highest difference (less than 50 million (mean value 2,45) and more than 500 million (mean value 3,59) is connected to the innovation method Big Data. These finding will be analyzed in detail in the following section.

        3.5 Specific Evaluation of Applied Innovation Methods The data set analysis shows deviations depending on the applied innovation method (see Fig.

         ] (statistic measurement of dis-

        tance to mean value) is owed to the disparity of the applied innovation method in each customer segment for the classes revenue and value chain. u u v ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi X N r t

        1 2 k ; l ¼ i ; l l ð2Þ x x N i



        r 170 K.-H. Lüke et al.

      k ; l : Standard deviation for each innovation method ð l ¼ 1 Þ and for each

      ; ; . . .; L

        customer segment class ð k ¼ 1 Þ, here: L ¼ 18 and K ¼ 2 (customer . . .; K segment class “revenue” (1) and customer segment class “value chain” (2)) ; x l : mean value of each innovation method l ð ¼ 1 Þ (over all segments) . . .; L : mean value of customer segment i for the innovation method l; for customer x i ; l segment class k

        ¼ 1 (“revenue”) N ¼ 3 (less than 50 million, 50–500 million, more than 500 million); for customer segment class k ¼ 2 (“value chain”)

        N ¼ 4 (OEM and Airline, Tier 1, Tier 2, Tier 3-n)


      Fig. 5. (a) Comparison customer segment “Revenue” (b) Comparison customer segment “Value


        It corresponds to the deviations of each customer segment (e.g. less than 50 million,

        and 50–500 million, more than 500 million) in the class (e.g. revenue). From Fig.

        (b) the following findings can be drawn:

      • For the customer segment class revenue, the innovation method Big Data has by far the highest value (rank 1) in relation to the standard deviation (0,52). The method “megatrend study” follows on rank 2 (0,43).
      • For the customer segment class value chain, the following four innovation methods are very similar in relation to the standard deviation: Scenario Engineering (0,35), Megatrend Study (0,34), Innovation Communities (0,33) and Big Data (0,33).

        Comparing the standard deviations of both customer segment classes on the basis of the collected data in the aviation industry, the Big Data innovation method seems to be valued differently. This implies that this method is variously applied within the customer segment classes. Big Data has huge potential for the generation of new ideas

        Innovation Management Methods in the Aviation Industry 171

      Fig. 6. (a) Ranking of standard deviation in the customer segment class “Revenue” (b) Ranking

      of standard deviation in the customer segment class “Value Chain”

        and innovations. Therefore, it should be considered in more detail. In the following section, the Big Data innovation method is introduced as an essential tool for the generation of new ideas in the aviation industry.

        4 Big Data Analytics as an Essential Method for Innovation Ideas Within the Aviation Industry

        4.1 Fundamental Aspects of Big Data The aviation industry can generate new ideas for innovations from Big Data analysis. This method offers opportunities for the aviation industry and its value chain, devel- opment, production, design and operations. The term Big Data can be defined and interpreted in various ways. Among these different definitions, the well-known three V´ s as in Volume (growth of data), Velocity (processing and analysis of data) and Variety (different structure of data) describe the required management of structured,

      • semi-structured and unstructured data [

        The enormous growth of data volume

        will be an exceptional challenge for the industry, nowadays and in the future. It is estimated that the worldwide data volume will be approximately 10 times higher in 2025 than in 2016 ].

        The critical issue for the aviation industry will be the extraction of information and knowledge from Big Data to increase both operational efficiency and innovations. The discovery of new ideas for innovations using special data analysis methods (Big Data Value Discovery) and the implementation of innovations (Big Data Value Realization)

        172 K.-H. Lüke et al.

        will play an important role in the future

         ]. When using Data Mining algorithms as a

        superior analytical method, new hypothesis and knowledge from the existing data set can be discovered. The Data Mining field covers several disciplines from mathematics,


        artificial intelligence, data management and database technologies [ In a data warehouse, structured data is stored in a relational database system. The aviation industry must manage data from various sources and of different structure such as flight data, maintenance data, fuel saving data, transportation data and sensor data from production systems all along the value chain

         ]. Traditional relational database

        systems reach their limits because of limited data volume, flexibility and processing

        4.2 General Technical Requirements Big Data technologies can help the aviation industry generating innovations. Therefore, the objective is to work out a concept for a general architecture. General requirements for this concept will be derived from market insight and technical aspects


      • OEM aircraft manufacturers and suppliers:

        Analysis of manufacturing data to suggest reliable materials, data exchange between suppliers and OEM concerning production, materials and supply chain data; Suggestions for design ideas concerning reliability, comfort, fuel savings and security can be made;

      • Operation improvements for airlines:

        Providing personalized services to the users; Forecasting possible faults of the aircraft; Forecasting possible deviations from scheduled flight data; –

      • Maintenance suggestions:

        Predictive Maintenance: using M2M (machine to machine) sensor data from aircraft and M2M data from production systems;

      • Routine Maintenance;

        Fault Repair;

      • Operation improvements for airports: using market data for flight information management
        • – improvement in flight monitoring;
        • – improvement in flight scheduling; Therefore, a general architecture concept is needed that covers the general technical requirements for an integrated Big Data solution, and special requirements for the generation of new ideas and innovations.

        4.3 Reference Architecture The proposed Big Data platform architecture (Fig.

        should acknowledge the gener-

        ation of new ideas for innovations and different sources of Maschine-to-Maschine (M2M) data. Flight data from airlines can be transmitted to the Big Data platform in

        Innovation Management Methods in the Aviation Industry 173

        real-time based on discrete time intervals or on demand during the flight. This data covers e.g. the position of an aircraft and other technical parameters. Whereas not all parameters are transmitted to the central Big Data platform during a flight, all relevant flight data parameters can be provided to the platform on the ground. Combining the respective data set with e.g. technical materials data or supply chain data from the manufacturers and suppliers, maintenance recommendations can be generated and sent to the maintenance center. Furthermore, the consolidation of the operational data (e.g. transport market data and time schedule data) from the airports, flight data from the airlines and user profile data (e.g. preferences) can give new insights for suggestions in flight arrangements [

        The proposed architecture of a Big Data platform is based on Big Data collection which is analyzed on different layers such as the processing layer, the Big Data ana- lytics layer and the Big Data knowledge layer


      • Big Data Collection and Processing Layer:

        Data Collection Module: this module collects the relevant data from different sources, it depends on the well-known ETL (Extraction, Transformation, Load)- Process in order to generate a common data format. Data Management: the collected and transformed data will be organized according to the data sources. Data Storage Management: the platform should provide different database technologies covering structured, semi-structured and unstructured data sets.

        A relational databases structure (e.g. SQL), and NoSQL technologies (e.g. MongoDB or Cassandra) should be supported by the platform. Service Management: this module covers the platform management including e.g. security, authentication, authorization services as well as performance and utility management.

      • Big Data Analytics Layer: the platform supports different algorithms from mathe-

        Data Mining Methods: matics and artificial intelligence to analyze data sets, increasing new insights for innovations. Computing Module: different computing frameworks should be supported according to the chosen algorithm, size of data sets and requirements of real time or non-real time analysis approach.

      • Big Data Knowledge Layer: this layer gives the opportunity to generate new ideas for innovations by providing different tools for visualization, by combining visu- alization tools with knowledge questionnaires to recognize new aspects and cor- relations. The two processes Big Data value discovery and Big Data value realization should mainly be supported ].

        Within the process step Big Data value discovery, business and innovation goals of an organization should be defined. It is important to consider several sources for innovations to balance the reasons whether innovations should be continued or stopped. The process step Big Data value realization covers Big Data design which deals with specific architecture elements, specific patterns and different technologies. Big Data Collection and Processing Layer

      1); 2); 3); 4); 5); 6)

      7); 8) 1) 2)

        Airlines Manufacturer and Supplier 3) Airports

        Maintenance Center 4) User-Profile- Data 5) M2M M2M M2M M2M M2M Dynamic Flight Data Airline Time Schedule Data Production and Material Data Supply Chain Data Data: Flight Monitoring and Scheduling Transport Market Data Data: Predictive Maintenance Data: Routine Maintenance Profile Data Personalized Service Offering

        Data Collection Module Maintenance Data Market Data Schedule/Flight Data Data Management Data Storage Management Relational Database NoSQL Database (e.g. Graph based, Document based) User Profile Data Production Data Supply Chain Data Service

        Management Security Authentication Authorization Big Data Analytics Layer Data Mining Methods Machine Learning Association Methods Artificial Intelligence Mathematics/Statistics Parallel Computing Grid Computing In-Memory Technologies Computing Module

        Big Data Knowledge Layer Visualisation Knowlegdge Questionnaire Hybrid Combination of Methods 6) 7) 8) Big Data Platform ETL (Extraction, Tranformation, Loading) - Process Big Data Value Discovery Big Data Value Realization Fig. 7. General big data architecture concept for the aviation industry 174 K.-H. Lüke et al.

        Innovation Management Methods in the Aviation Industry 175

        5 Summary and Outlook According to the results of the empirical survey in Sect. , eighteen different innova-

        tion management methods are introduced and their application level in the aviation industry is examined. The methods are assigned to the cooperation fields: supplier, customer and in-house. Although the innovation methods “megatrend study”, “test market” and “external contest for ideas” are currently most frequently applied at all customer segments, there are remarkable differences in the application level of the different customer segments. Comparing the standard deviations for the customer segment classes revenue and value chain, the investigated data set clearly qualifies the Big Data method as highly promising. Big Data has huge potential for the generation of new ideas and innovations.

        Based on general technical requirements, a general Big Data architecture concept for the aviation industry is introduced in Sect.

         . Combining different data sources from

        airlines, manufactures and suppliers, airports and maintenance centers and user profiles, a layer-architecture for the aviation industry is described. The proposed architecture concept should be discussed with representatives of the aviation industry. Furthermore, a comparison between different industry sectors, e.g. automotive industry, would be helpful to work out similarities and dissimilarity in the application of innovation methods.



      1. Richter, K., Witt, N.: Introduction. In: Richter, K., Walther, J. (eds.) Supply Chain

      Integration Challenges in Commercial Aerospace. Springer, Cham (2017).



      2. Deloitte: Global Aerospace and Defence Sector Financial Performance Study (2016).

      . Accessed Aug 2016


      3. KPMG: The Future of Civil Aerospace, KPMG UK (2013).

      Accessed May (2016)


      4. PRN: Global Aerospace and Defence Industry to Grow Around 3%, PR Newswire, 12

      February 2015

        5. Airbus Annual Report (2016). . Accessed 22 Feb 2018


      6. Crossan, M.M., Apaydin, M.: A multi-dimensional framework of organizational innovation:

      a systematic review of the literature. J. Manag. Stud. 47, 1154–1191 (2010)

        7. BCG: Most innovative Companies, 17 January 2018. Accessed 22 Feb 2018


      8. Shenhar, A.J., Holzman, V., Melamed, B., Zhao, Y.: The challenge of innovation in highly

      complex projects: what can we learn from Boeing’s dreamliner experience. Proj. Manag.

        J. 47, 62–78 (2016)

      9. CG: The Changing Face of the Aerospace and Defence Industry, Capgemini, Paris (2011).

      . Accessed 22 Feb 2018

      10. The Boeing Company Annual Report (2016).

         Accessed 22 Feb 2018


        176 K.-H. Lüke et al.


      11. Eßig, M., Glas, A.H., Wäldchen, D.: WP25 – Innovation Exchange in Supplier-Buyer Dyads



      12. Myers, S., Marquis, D.G.: Successful Industrial Innovations: A Study of Factors Unterlying

      Innovation in Selected Firms. National Science Foundation (1969)


      13. Ili, S., Albers, A., Miller, S.: Open innovation in the automotive industry. In: R&D

      Management 4.0 (2010)


      14. Walther, J., Wäldchen, D.: Open innovation in the aviation sector. In: Richter, K., Walther,

      J. (eds.) Supply Chain Integration Challenges in Commercial Aerospace, pp. 57–71.

        Springer, Cham (2017).


      15. Adamczyk, S., Bullinger, A.C., Möslein, K.: Innovation contests: a review, classification and

      outlook. Creat. Innov. Manag. 21(4), 335–336 (2012)


      16. Dahlander, L., O’Mahony, S.: A study shows how to find new ideas inside and outside the

      company (2017). Accessed 22 Feb 2018


      17. Bullinger, H.-J., Warschat, J.: Innovationmanagement in Netzwerken. In: Garcia Sanz, F.J.,

      Semmler, K., Walther, J. (eds.) Die Automobilindustrie auf dem Weg zur globalen Netzwerkkompetenz, pp. 199–214. Springer, Heidelberg (2007).


      18. Fichter, K., Beucker, F.: Innovation Communities, Promotorennetzwerke als Erfolgsfaktor

      bei radikalen Innovationen, Stuttgart (2008). Accessed 22 Feb 2018


      19. Schmolze, R.: Unternehmen Idee, Wie kundenorientierte Produktentwicklung zum Erfolg

      führt, Frankfurt/New York (2011)


      20. Blattberg, R., Golanty, J.: Tracker: an early test market forecasting and diagnostic model for

      new product planning. Perspect. Promot. Database 177–187 (2010). . Accessed 22 Feb 2018


      21. Katsikis, N., Lang, A., Debreceny, C.: Evaluation of open innovation in B2B from company

      culture perspective. J. Technol. Manag. Innov. 32(3), 95–100 (2016)


      22. Breuer, H., Steinhoff, F., Wogatzky, M.: User Clinic Formate und ihr Beitrag zur

      Innovationserfolgsrechnung. In: Schmeisser, W. (Hrsg.): Technologiemanagement und Innovationserfolgsrechnung, S, pp. 1–23 (2010)


      23. Steinhoff, F.: Tools for user-driven innovation at Deutsche Telekom Laboratories. In:

      Arnold, H., Erner, M., Möckel, P., Schläffer, C. (eds.) Applied Technology and Innovation Management, pp. 72–88. Springer, Heidelberg (2010).


      24. Chapodos, J., Desmond, K., Schlett, Ch.: Creating a suggestion system as part of a continous

      improvement initiative (2014). . Accessed

      22 Feb 2018 25. Höber, B.: Theoretical background: internal innovation contests and the work environment.

        In: Höber, B. (ed.) Firm Internal Innovation Contests. Work Environment Perceptions and Employees’ Participation, pp. 17–59. Springer, Wiesbaden (2017). . Accessed 22 Feb 2018

      26. Barbour, M.: How to run an ideation workshop that actually leads to innovation (2016).

        Accessed 22 Feb 2018


      27. Durance, P., Godet, M.: Sceanrio building: uses and abuses. Technol. Forecast. Soc. Change

      77, 1488–1492 (2010)

        Innovation Management Methods in the Aviation Industry 177


      28. Oliveira, M.G., Rozenfeld, H.: Integrating technology roadmapping and portfolio manage-

      ment at the front-end of new product development 77(8), 1339–1354 (2010)


      29. UN Global Pulse: Big Data for Development: Challenges & Opportunities, May 2012. Accessed Feb 2018

      30. Lüke, K.-H., Kapitány, D.: Business (lead) customer involvement in the innovation process.

        In: Arnold, H., Erner, M., Möckel, P., Schläffer, C. (eds.) Applied Technology- and Innovation Management, pp. 59–71. Springer, Heidelberg (2010).


      31. Rohrbeck, R., Thom, N., Arnold, H.: IT tools for foresight: the integrated insight and

      response system of deutsche telekom innovation laboratories. Technol. Forecast. Soc.

        Change 97, 115–126 (2015)


      32. Backhaus, K., et al.: Multivariate Analysemethoden. Springer, Heidelberg (2016).

        33. Bleymüller, J.: Statistik für Wirtschaftswissenschaftler, München (2015)


      34. Chen, H.-M., et al.: Towards Big Data value engineering for innovation. In: Proceedings of

      the 2nd International Workshop on Big Data Software Engineering, Austin, USA, pp. 44–50.

        ACM (2016)


      35. Singh, A., Kaushik, A.: Knowledge based retrieval scheme from Big Data for aviation

      industry. In: IEEE International Conference on Computational Intelligence and Communi- cation Networks, pp. 918–923 (2015)


      36. Müller, S., Keller, C.: Pentaho und Jedox, Business Intelligence-Lösungen: Data

      Warehousing, Reporting, Analyse, Planung. Hanser-Verlag (2014)


      37. Statista 2017: Prognose zum Volumen der jährlich generierten digitalen Datenmenge

      weltweit in den Jahren 2016 und 2025 (in Zettabyte). IDC. . Accessed

      8 Feb 2018


      38. Runkler, Th.: Data Mining, Methoden und Algorithmen intelligenter Datenanalyse,

      Wiesbaden (2015)


      39. Li, S., et al.: Civil aircraft Big Data platform. In: IEEE 11th International Conference on

      Semantic Computing, pp. 328–333 (2017)


      40. Cuzzocrea, A., Loia, V., Tommasetti, A.: Big-data-driven innovation for enterprises:

      innovative big value paradigms for next-generation digital ecosystems. In: Proceedings of Seventh ACM International Conference on Web Intelligence, Mining and Semantics, Amantea, Italy (2017)


      41. Ayhan, S., et al.: Predictive analytics with surveillance data. In: Proceedings of

      ACM SIGSPATIAL BIGSPATIAL 2012, pp. 81–90 (2012)


      Digital Transformation in Companies –

      Challenges and Success Factors

        Marcus Wolf

        ( B


        , Arlett Semm, and Christian Erfurth


      Department Industrial Engineering, University of Applied Sciences Jena,

      Carl-Zeiss-Promenade 2, 07745 Jena, Germany



        New supply networks, more automation, self-regulating

      products - the opportunities of digital transformation seem infinite. Some

      companies are facing very solid issues when it comes to implementing

      Industry 4.0. As part of the research project “Healthy Work in Pioneer

      Industries”, case studies were conducted with companies and reflected

      on the current state of research. It can be said that companies have fun-

      damental challenges in mastering the digital transformation: Isolatory

      thinking, no active knowledge management, underestimation of digital-

      isation and lack of knowledge, no resources and no awareness of digi-

      talisation tended to be identified as deficits. There are also examples of

      how companies are successfully addressing digital transformation. The

      identified success factors are the creation of innovation areas, network-

      ing across company boundaries, the application of agile methods, the

      motivation to try out new things and the active management of digital

      transformation in the company. In this paper, we discuss obstacle and

      success factors we observed in our case studies on basis of a cross case

      study analysis.

        Keywords: Digital transformation


        Industry 4.0 ·

        Case studies

      1 Introduction

        The introduction of technological innovations drives many companies to cope with more challenging demands from customers. With the digital turn, the search for holistic approaches was started by companies to change successfully and to stay on the market. There raise some questions regarding the way we work in the future and connected with this ways of changes. As part of the research project “Healthy Work in Pioneer Industries” which focuses on case studies in the field the research team of the University of Applied Science Jena is inves- tigating aspects of introducing technological innovations in companies. Results of this work will be presented in a case study archive. It contains case studies developed by the interdisciplinary collaboration with scientists from psychology, medicine and social science within the project. The reason for this is that it is unclear how the change through digital transformation affects healthy working.

        Digital Transformation in Companies – Challenges and Success Factors 179

        New reliefs and burdens are created. For this purpose, inspections and exploratory interviews are carried out for the creation of individual case studies. First results, which issues companies have to overcome when introducing new technologies, which success factors can be identified and what are the driving technologies, are presented in this paper.

        A major influencing factor for the need of changes is the digital transfor- mation actual happens. In our study, we have tried to look at the company holistically, because digital transformation means a permanent change that can affect various aspects of the company. We have focused on the following aspects of digital transformation:

      • – Drivers of digitalizations
      • – Objectives behind digitalization
      • – Structural challenges
      • – Cultural challenges
      • – Data – Effects on work
      • – Interorganizational Effects – Resources for digitalization

        The digital transformation is not only characterized by new technologies, but also in particular by the combination of new and old: evolution and revolution. Existing technologies can be combined with new technical possibilities. We have focused our investigation on the following technology areas:

        1. Augmented Reality

        2. Artificial Intelligence (AI)

        3. Big Data

        4. Blockchain

        5. Internet of Things (IOT) 6. 3D printing

        In comparison to current research literature, we have identified other fun- damental issues. This leads to the research question of how a successful digital transformation of companies can succeed in the face of current challenges. This paper discusses initial approaches.

      2 Related Works

        In recent years more and more research activities are focused on industrial IoT and digitalization in industrial contexts. There are some case and empirical stud- ies on different topics. For instance Groggert et al.

        investigate status quo and

        potentials of manufacturing data analytics. Mostly data is used for predictive maintenance and process optimizations. In many cases data is collected but not further used often. This is also stated in

         ] where the authors defined a maturity

        model for digital readiness. With data representing a digital twin of a product the

        180 M. Wolf et al.

        knowledge of state and identity of the product can be used in different scenarios. E.g. in

         ] a digital twin-based approach for rapid individualized designing of the hollow glass production line is investigated.

        Approaches using augmented reality can be found in combination with assis- tance systems. In

        the authors give insight into an research project examining

        the potential of smart glasses used as a component of assistant systems for work- ers performing maintenance tasks in an industry 4.0 context.

        Since fast delivery in a high quality is essential for suppliers digitalization is a promising way to meet the demands. In

         ] an in-depth case study of one

        textile - apparel industry in Vietnam is given. An analytical study of conditions for adopting the horizontally collaborative transportation strategy in competi- tive circumstances is shown in ].

        Known from software development agile methods are more and more common even in the area of production. The paper

        focuses on challenges in the

        implementation. First steps in the direction of a holistic agile process framework are shown in


        The role of workers and employers in the process of digitalization attracts attention. The success of a digital turn is gained when acceptance of technology and process changes is achievable. One influence factor seems to be the age of the workforces

         ]. The productivity of workers is also one of the research


        . However processes get more complex and future work is changing. In



      ] explorations are made on basis of interviews. More holistic research can be

      found (challenges, roadmaps, knowledge management).

      3 Conduction of Cross Case Study

        The study was conducted in the form of exploratory case studies. For this pur- pose, a standardized questionnaire with guiding but open questions was cre- ated in the research project. Some fundamental questions were generally raised. Supplementary questions served to deepen some facts if necessary. The conversa- tional partners could talk freely. In-depth questions were asked depending on the situation and answers. The interviewees came from different areas of the compa- nies, e.g. IT manager, HR department, employees from specialist departments. It was possible to ask specific questions about technologies or methodologies and to get answers from different view points.

        The questions in the guide are assigned to different topics: Technological changes in recent years are recorded, associated changes in work, the workloads of employees and the effects on occupational health and safety, as you can see in Fig.


        The methods were applied in the research project to identify new correlations, problems and solutions in relation to the Digital Transformation. Various persons in the company were interviewed for this reason. However, a guideline-based expert interview could not always be conducted. At the same time, the reasons can also be seen as major methodological problems: (1) The expert interview based on guidelines could not yet be carried out in the current research project.

        Digital Transformation in Companies – Challenges and Success Factors 181

      Access Industry profile Basic data


      Organizational Application of Health burden

      structures new technologies and relief

      Occupational health and safety,

        Human Resources Perspectives, health Management Strategies management, health promotion Fig. 1.

        Overview of topics, guideline questions and case study structure

        (2) The relevant target groups for the guideline-based expert interview have not yet been reached. In Fig.

         you can see the main questions.

        In addition to the interviews based on the guidelines, the case studies also included operational documents, protocols from inspections and discussions with the companies for consideration.

        We have contacted various companies and received information. After an initial access to the companies we had difficulties to get more detailed informa- tion. Nevertheless, there were some companies that gave us deep insights. This has resulted in case studies and cover various industries: mobility, production, mechanical engineering or e-commerce.

        182 M. Wolf et al.

        D. Decentralization/Service orientation

        The internal drivers for digitization can be based on an expected increase in efficiency or a new innovation. These types of drivers are coming out of the company. The companies want to improve existing products or services or add new features.

        4.1 Issue: Drivers of Digitalizations Observation: We can differentiate between external and internal factors in the case studies. For the companies there are only internal factors, only external factors or a mixture of both categories that work in relation to the Digital Trans- formation.

        With the application of our case studies we could observe:

        In the following we present the results of our case studies in relation to Digi- tal Transformation. For this purpose we have developed theses, which we also explain in this chapter. We have also assigned these theses to topics which - in our view - can be identified as key factors influencing the Digital Transformation of companies. In the following chapter we will discuss the findings.

        Overview of detailed guideline questions and case study structure for applica- tion of new technology. The technological functional areas are based on p. 12].

        E. Self-organized processes Proceeding to introducing or testing of technological changes Fig. 2.

        C. Networking/Integration


      Application of new technologies

        B. Assistance systems

        A. Data acquisition/processing

        Main reasons for introducing Classify of the technologies (T1, T2, T3 etc.):

        1. Technological inventory and organization Modifications or renewals in recent years Most technical changes (only significant technology: T1, T2, T3 etc.)

        IT-Solutions Decentralized use of IT-Systems: Effects of processes in the company Interaction between mechanics, electronics and software with new technologys Automatic control in the use of technologies Realization of automation from the technology

        2. Technological functional areas Kind of collected data Use of the data Description of used technology offers assistance functions Mistakes and limits of application External collaboration with partners with

      4 First Findings from Our Case Studies

        Digital Transformation in Companies – Challenges and Success Factors 183

        Besides the internal drivers there are also the external drivers. These are, however, located outside the company. They are characterised by a different type of customer expectation, specifications from the manufacturer or supplier and from regulatory authorities.

        Observation: The drivers of digitization can be promoted in the company by the management or from the departments and especially from the IT department.

        The Digital Transformation has an effect on the whole company. Depending on the company situation, different areas of the company are focusing on digiti- zation. Management may be aware to worry about the existing business model, which may be endangered by digitization. The specialist department may have received customer requests that require an innovative or more efficient product or service design. A company’s IT department knows the new technological and

        IT opportunities. They can proactive go with new ideas to the departments and management. Observation: Companies react proactively or reactive.

        The change to Digital Transformation in companies is being driven by compe- tition, new technologies, customer demands and regulatory requirements. Behind this is always the economic benefit, which is decisive for the companies.

        It has been shown that there are proactive companies as well as reactive companies. In a case study, for example, start-ups were specifically searched in the context of shaping the Digital Transformation with external competencies. Other companies are reactive in relation to Digital Transformation. Especially requirements and dependencies on suppliers and manufacturers as well as on regulatory and supervisory authorities seem to shape technological development in companies.

        4.2 Issue: Objectives Behind Digitalization Observation: There are different motivations and expectations to be found in the management, executives and employees in relation to the Digital Transformation.

        The Digital Transformation has an effect on a wide variety of business areas. Each area and each position has its own motivation. 1. The management has the goal of conserving the existing business model under the influence of Digital Transformation and opening up new business models and thus business areas.

        2. The employees have a tendency to secure the job or to get work facilitation regarding the accomplishment of routine tasks and facilitation in the form of supporting technologies in the accomplishment of complex tasks. 3. The execu- tive employees have the intention that the responsible department will perform its tasks for the success of the company. This can be endangered by the Digital Transformation, so that there is particular interest in further developing the area for the future.

        In the fashion retailer case study, the management was the driving force behind the Digital Transformation. The problem here is that the measures do not come from the employees and can therefore be expected to be difficult to

        184 M. Wolf et al.

        implement. This is particularly about the acceptance problems. The above expla- nations make it clear that there are different motivation levels in the companies. These should at best be used in relation to Digital Transformation, because 1. the Digital Transformation has an impact on the entire company, in other words it affects different areas and 2. the inclusion of the various motivations can generate an benefit for all involved. The fashion retailer has been strictly communicated with the employee representatives. Previously, every topic relating to Digital Transformation was closely coordinated with the management. These internal expenditures hinder the innovative power to the outside. Therefore, communi- cation should be transparent and the digital transformation should be designed together with the participants.

        Observation: From a macro-perspective it can be seen that companies implement measures to increase innovation or efficiency.

        We were able to find both target measures in the companies. The two target measures show themselves differently in the companies. For example, the effi- ciency can be seen in the fact that the throughput speeds want to increase. In the case study of mechanical engineering we have observed that the cleaning of parts should be driven forward by an IoT project. Concentration measurements should be carried out with the aid of sensors in order to optimise the cleaning of components in corresponding tanks and at best to avoid double cleaning pro- cesses by improving the concentration ratio of the cleaning liquid. The goal of this project is to improve the process and increase its effectiveness. In addition to this aspect, increasing efficiency in companies can also result in more cost- effective work (e.g. less consumables) or faster work (e.g. greater utilization of existing production facilities and resources).

        Innovation as motivation for the company can also be seen in various points. It can be that the quality of a product or service is to be optimized by new technologies. It can also be that the purpose of innovation is to make itself more independent of existing value network partners or to establish a new busi- ness field with technological innovation. In addition, technological innovations offer the potential to respond in a special way to the demands and pressures of employees. In the case study of automotive suppliers, the goal was to use new technological solutions to make it easier to adapt the workplace to the respective employee.

        4.3 Issue: Structural Challenges Observation: Established big structures in and around the companies are tending to hinder.

        Complicated administrative apparatuses that ensure the operational busi- ness traditionally hinder innovation. In addition, Digital Transformation requires management and the specialist department. Many people in the company have to be involved. Here one can speak of the disruptive side of digital transformation, because these responsibilities and processes are changed. Innovation requires flat structures, short paths and a free space to realize ideas. Highly regulated struc- tures prevent digitization. Start-ups have it easier there, because most of them

        Digital Transformation in Companies – Challenges and Success Factors 185

        have not yet established a complicated organizational apparatus. Therefore, it is also easier for small companies to transform digitally, because they usually have simpler structures: Small business structures have flat decision-making lev- els which promote creativity and the testing of new technological solutions. In addition, decision-makers are closer to digitization issues. Observation: Agile methods can support.

        The used methods in the structural units are also essential: 1. learning from mistakes, 2. a quick evaluation of development paths (with regard to the com- plexity of technical solutions and possibilities) and 3. a quick trial and error should support the methods. Agility seems to be the right tool of choice. Meth- ods like Kanban can help teams to transparently support where they stand with the processing of tasks and where support is necessary. Scrum as a method allows a review after short sprints, e.g. three weeks: What did we want and where are we now? An adaptation of the development can happen accordingly fast. The important thing here is that all this requires a structure that allows it. Observation: In information technology there is a conflict of goals in relation to the necessity of change and the guarantee of safe operation.

        On the one side, a company always has the intention to offer its products and services reliably on the market, usually with a higher quality. Interventions that become necessary as a result of innovation and changes in technologies can disrupt the safe operation. Different concepts, e.g. “Bimodal” or “DevOps” can help to resolve this target conflict. Observation: The lack of a responsible area for the digital transformation of the company can have a negative impact on future business development.

        A position or area is required to manage the digital transformation of the company for the following reasons: 1. The Digital Transformation impacts on the whole company. It requires knowledge of the structures, resources and offers of the company to make this digitization possible.

        4.4 Issue: Cultural Challenges Observation: Diversification of employees: Bringing together wisdom and enthusiasm

        Digital Transformation also seems to be a generational question. We have seen a tendency for older workers to have problems understanding the conse- quences of Digital Transformation. Younger workers seem more open to new technologies. For this reason, companies should promote diversification in gen- eral and in particular with age. The goal is to bring together access to resources and the wisdom of older workers with the enthusiasm and experience of digi- tal technologies with younger workers. It can also be useful to actively involve younger people in management.

        This also means that the path via disruption is limited. Doing everything new, i.e. without wisdom and access to resources is not possible. In this context, the term disruption does not seem to be appropriate in relation to a principle of action with a view to digital transformation.

        186 M. Wolf et al.

        Observation: Enhancement of knowledge exchange: If a company is more frag- mented with existing knowledge, it is less successful.

        A fragmented knowledge in the company, knowledge which is only accessible to a part of persons, prevents goal-oriented evaluations and related measures. In the more complex world of work, it is necessary that the departments of a company and the people exchange information as transparently as possible in order to clarify company situations, product and service expectations. There is a tendency for products and services to be increasingly considered in their full life cycle in view of better satisfying customer expectations. In addition, Digital Transformation covers a company as a unit. It is therefore essential to promote a transparent exchange of knowledge.

        An essential factor is also the binding of knowledge to the company: Expe- riences, solutions of problems are often highly person-dependent. In an increas- ingly specialized work environment, a loss of know-how is increasingly a critical problem for companies. This can be counteracted by actively managing feedback and knowledge in companies. Knowledge must be tied to the company as far as possible.

        One challenge, is to promote this exchange of knowledge. Employees have less motivation to share knowledge from their position, because there is a danger that they will become replaceable. The advantage of knowledge transfer is that it can also be an opportunity to constantly open up new fields and thus continue to hold a significant position in the company or the job market. Observation: Lack of openness to collorabotion endangers the digital transforma- tion of companies.

        There can be several reasons for a lack of collaboration culture in companies. It is possible that the company has inadequate incentive systems to prevent employees from working well with others. This can also be attributed to a lack of tolerance, which can also be established as a corporate goal. People are wor- ried about their creative spaces, which they sometimes have under own control.

        Colloaboration always means a modification of the own sphere of influence. How- ever, it is important in terms of Digital Transformation in order to work together to further develop the products and services for the company’s success. Observation: Digitalization needs constant changes.

        The change affects people in different ways. New conditions are being estab- lished, new structures and a new form of management are needed. The techno- logical changes have to be widely accepted in the company: the employees, the managers, the organs of a company.

        This is particularly problematic in view of the established routines of humans. People love new things, but don’t like change. This can be seen in particular in various legal frameworks that make certain formal requirements or regularisation by regulatory authorities, which requirements a business process has to fulfil.

        However, constant change has to be established within the company in order to remain competitive.

        Digital Transformation in Companies – Challenges and Success Factors 187

        4.5 Issue: Data Observation: Intransparency in data sharing prevents the creation of new value- added networks and endangers business success.

        Everyone wants to have data accessible, nobody wants to release data with- out motivation. We also see typical human behavior in companies in relation to value chains. Targeted sharing of data in a value-added network can promote adaptation to market conditions and thus promote the own Digital Transforma- tion. Companies often have the intention, as we have seen in our case study of the railway company, not to disclose data to competitors, but to share it with the manufacturer or supplier. These data were assessed as decisive for compe- tition. This assessment, including data protection problems, contrasts with the progress of the value network and thus also indirectly with the existence of the company. Therefore, data should be exchanged in a target-oriented way.

        Observation: Value creation only works optimally if data can be merged.

        In order to break up existing business practices, it is necessary to have access not only to the available knowledge, but also to the data available in the com- pany. The knowledge of company data is just as necessary for the Digital Trans- formation as the active management of knowledge. Who are our customers, what is selling well and how will this change the digital transformation? These ques- tions can help to align the business field.

        Data silos also inhibit productivity. Who has access to what information, how do I get access, etc. - this creates an administrative overhead. In addition, this often results in multiple efforts in communication if certain facts or goals have to be communicated to different groups of people again and again. This is why the flow of data within the company should be as free as possible.

        4.6 Issue: Effects on Work Observation: Digital Transformation creates a new risk and opportunity in terms of differentiating the level of performance and demands of work.

        The level of performance and requirements can change as a result of Digital Transformation. On the one side, the new technologies offer opportunities to respond to the different demands and performance levels of employees, because the work is more routine, especially as a result of automation. Error reduction, no rework and relief from monotonous activities are the positive consequences.

        On the other side, the complexity tends to increase with certain activities and the spontaneously demanded possibility of intervention of the employee. There are therefore risks with regard to the monotisation of work through automation (if not equal elimination of work) as well as an increased spontaneous high sense of decision and responsibility. It obliges us, the human being, how we use the technological tools.

        In our case studies we could see the following example. In the case study of fashion traders, it appeared that two different employee groups were formed in the company. On the one side the cognitive resilient and capable IT employees,

        188 M. Wolf et al.

        who try to further develop the company by increasing efficiency and innovation. In addition, these employees have various advantages, for example an employee who cares about the interests of this group of employees. On the other side, we have seen simple logistics staff carrying out typical logistics tasks within a fixed timeframe. You are monitored by the number of activities performed. So while one group can be creative and no direct monitoring of the activity is carried out, the other logistics group does not encourage creativity, but monitors it. At the same time, logistics employees are sometimes used as test users without their knowledge.

        4.7 Issue: Interorganizational Effects Observation: The dependency of partners can be inhibited.

        Within the market and value creation network there are dependencies for companies. In particular these, which exist with current partners, can provide a background for own digitization efforts. This can be seen, for example, in the involvement of various partners in a business process. If the process is changed due to a change in technology, this must also be implemented by the corre- sponding partners. This results in higher communication efforts and provision of resources for the partner. The goal must be transparent for all involved partners.

        One way of best modifying a business process in a value-added network can come from the company or institution (public authority, etc.) which has the widest scope and supports the Digital Transformation here, for example, through financial incentives. It is also possible that a manufacturer of a product and a provider of a service may promote the introduction of a new technology by improving the quality or simplification of a product.

        Observation: Big companies dominate small companies.

        A particular problem is the market power of established companies in the associated value chains. Here, large companies tend to dominate small compa- nies. This means that small companies are more likely to be successful in the market with innovations and efficiency improvements in specific areas. Start- ups in particular have the potential to use the assets of established companies for their products and services. Large companies can use their market power to enforce technological specifications for the service or product they use.

        4.8 Issue: Resources for Digitalization Observation: Digitization requires resources.

        Digital Transformation can only succeed in companies if adequate resources are made available for it. The reasons for this can be found in the complexity and diversity of the new opportunities of action. You have to try it out in the short time, if you want to be sure of the existence of the company in the long time. This includes both financial aspects as well as employees from different areas (including management) of the work. Resources should be chosen based on the specific purpose, market environment and business objective in relation

        Digital Transformation in Companies – Challenges and Success Factors 189

        to Digital Transformation. For many companies this means a lot of effort, as we have also observed in our case studies. Management must make the path possible (enablers who clarify the structural question and set corporate goals), the employees try out new pioneering ideas and technologies and then try to establish them if necessary. This also often means, in contrast to habits, trying out new things. However, sooner or later the companies, as can be generally observed, will be forced into action by new “players”.

        Companies can economize resources by using existing standard technological solutions. As a rule, it is not necessary to develop something new, what can save resources. Its better to adapt an existing framework, IT system or technology. Observation: Digitisation can be bought in to a limited extent in relation to resources.

        It is possible to purchase resources for the digital transformation. In partic- ular, we observed in our case studies that external know-how has been acquired. This can take the form of joint cooperation within the project or, as in busi- ness life always be visible, by buying up start-ups and specialised companies. Although cooperation with start-ups can be problematic in the long term (will the company establish itself on the market and continue to be available as a partner?), the goal is to realize projects together in the short term. The active operation of knowledge management is essential here, if you work together in such a form.

        In addition, however, it is essential to have located Digital Transformation as a position in the company. A position or team that coordinates these activities and is well aware of the corporate goals that enable Digital Transformation by linking the relevant areas in the company with external parties. This position must also be able to say where appropriate resources need to be mobilised in the company.

      5 Discussion

        We find out different positions from the companies to manage the Digital Trans- formation. As a result from the research literature and our case studies we iden- tified this possibilities to overcome obstacles. In Figs.

         you can see the relation between obstacles and possibilities.

        1. Drivers of digitalizations No matter if you are driven by external or internal drivers: It is essential that the Digital Transformation should be designed proactively in the context of the company’s scope of design and strategic orientation. If the company misses measures too late, there is a risk that the business model could break away.

        In addition, the new technological possibilities give us the opportunity to actively position the company in places that could create new value creation networks.

        2. Objectives behind digitalization There are different levels of motivation in relation to digitization. It is impor- tant that the interests of management - in particular to save unit labour

        190 M. Wolf et al.

        Internal: efficiency, Monotonous innovation workforce, External: customer fragmented

      expectations, Differentiated Existing structures knowledge, holding

      specifications expectations problematic on to the habitual Drivers of Objectives behind Structural Cultural challenges digitalizations digitalization challenges

      Proactive action, Relief of activities, Management has to Diversification of the

      securing / developing support with complex create the workforce, promotion

      business model tasks preconditions, teams of knowledge with agile methods exchange, stabilization of change

        Fig. 3.

        Obstacles and possibilities for drivers of digitalizations, objectives behind dig- italization, structural challenges and cultural challenges Small resources for digitization, lack of awareness of the impact of Digital

        Intransparency and Dependence on Transformation, gap limited/no shared partners and between day-to-day data storage inside Monotonous and authorities, market business and and outside the sometimes complex power of large innovation company situations companies

        Interorganizational Resources for Data Effects on work Effects digitalization Promoting the Adaptation of work Networking beyond Creation of targeted exchange of to different levels of industry boundaries innovation areas, data inside and performance and on specific topics, involvement of demands of outside the company make new challenges strategic partners employees, transparent communication of objectives


      Fig. 4. Obstacles and possibilities for data, effects on work, interorganizational effects

      and resources for digitalization

        costs - should be regulated by legislators. Through automation and tax advan- tages, through the use of technology and the reduction of previous human work, it is necessary to intervene in regulation in view of allowing as many people as possible to participate on the benefits of this progress. In addition, regulatory authorities should regulate the dangers of artificial intelligence at an early point in time so that no dangers can develop here. Entrepreneurs

        Digital Transformation in Companies – Challenges and Success Factors 191

        can also use the digital transformation to relieve people of boring work and to respond to the different demands and performance levels. It is also important to combine the different motivational positions in the company in view of successfully mastering the Digital Transformation. Everyone should be able to recognize an extra benefit in the transformation.

        3. Structural challenges The management of the company has to set the preconditions for successful digitization. Small agile teams need innovation spaces to try out new ideas.

        The Digital Transformation has to be organizationally located in the com- pany. Management has to be enablers. The difficulty here can also be seen in the continued successful management of day-to-day business.

        4. Cultural challenges A corporate culture should be promoted which supports diversification of employees, especially in view of age differences.

        5. Data Companies should promote data exchange with other companies in the value- added network and actively decide on the sharing of data depending on the use case. Data streams should become visible in the company. This implies the use of common information systems moving away from paper and indi- vidual spreadsheet solutions. Only if the data is available transparently in the company, it can be worked with the data and the company can be optimized.

        6. Effects on work Work can be designed to be more human: The different demands of employees can also be addressed. It is also essential to communicate the changes so that employees can identify with the goals of the Digital Transformation and recognize an benefit for themselves and a own perspective.

        7. Interorganizational Effects The Digital transformation is a complex process. In order to orientate oneself and find more efficient partners for new challenges, it can also make sense to network beyond the established industry borders. The networking with start-ups and other companies and science can be helpful in solving similar problems. In addition, the integration of the own company into a value net- work should be checked. In same way Furthermore, other companies should also be observed to find out if they can develop into a new unknown competi- tor as a result of the new technological possibilities.

        8. Resources for digitalization Digitization requires resources. These have to be created in the company and partly bought in.

      6 Conclusion

        The paper showed that Digital Transformation impacts different areas of the company. We have found that some companies have to solve very down-to-earth problems before further steps can be taken in direction to the Digital Transfor- mation. We have demonstrated the main obstacles on the basis of topics that

        192 M. Wolf et al.

        have been formed. We have tried to compare these obstacles with success factors in the context of good practise case studies and the latest research. The case stud- ies are generalized. They respectively the findings have to be classified according to the company situation. In summary it can be said that successful companies are characterized by 1. a change of mentality in the management, 2. creating the preconditions for an innovative area and 3. promoting the exchange of knowl- edge and data within the company. In addition, it remains to be researched which obstacles and success factors have an effect additionally. In the future, this can also be used to develop a model in which companies can see a guideline for solving fundamental challenges in relation to the Digital Transformation.


      Acknowledgment. The authors are very thankful for the support of our case studies

      specifically Andre Bahnemann, Jonas Katzenmaier and Laura Thiele. This work was

      supported in part by the German Federal Ministry of Education and Research (BMBF)

      under the grant number 02L14A073.



      1. Agiplan GmbH, Fraunhofer IML, ZENIT GmbH: Studie erschließen der potenziale

      der anwendung von ‘industrie 4.0’ im mittelstand [study on unleashing the potential of application of industry 4.0 among small and medium-sized enter- prises] (2015).

        Accessed 17 March 2018


      2. B¨ ohmer, A.I., Hugger, P., Lindemann, U.: Scrum within hardware development

      insights of the application of scrum for the development of a passive exoskele- ton. In: 2017 International Conference on Engineering, Technology and Innovation (ICE/ITMC), pp. 790–798, June 2017.


      3. Carolis, A.D., Macchi, M., Negri, E., Terzi, S.: Guiding manufacturing companies

      towards digitalization a methodology for supporting manufacturing companies in defining their digitalization roadmap. In: 2017 International Conference on Engi- neering, Technology and Innovation (ICE/ITMC), pp. 487–495, June 2017.


      4. Gleeson, F., Goodman, L., Hargaden, V., Coughlan, P.: Improving worker produc-

      tivity in advanced manufacturing environments. In: 2017 International Conference on Engineering, Technology and Innovation (ICE/ITMC), pp. 297–304, June 2017.


      5. Goevert, K., G¨ okdemir, A., Peitz, C., Lindemann, U.: Challenges of agile devel-

      opment implementation in mechatronic development processes. In: 2017 IEEE International Conference on Industrial Engineering and Engineering Manage- ment (IEEM), pp. 701–705, December 2017.



      6. Groggert, S., Wenking, M., Schmitt, R.H., Friedli, T.: Status quo and future poten-

      tial of manufacturing data analytics - an empirical study. In: 2017 IEEE Interna- tional Conference on Industrial Engineering and Engineering Management (IEEM), pp. 779–783, December 2017.

        Digital Transformation in Companies – Challenges and Success Factors 193


      7. Ho, D., Kumar, A., Shiwakoti, N.: Supply chain collaboration - a case study of

      textile and apparel industry. In: 2017 IEEE International Conference on Indus- trial Engineering and Engineering Management (IEEM), pp. 1367–1371, December 2017.


      8. Hobert, S., Schumann, M.: Enabling the adoption of wearable computers in enter-

      prises - results of analyzing influencing factors and challenges in the industrial sector. In: 50th Hawaii International Conference on System Sciences, HICSS 2017, Hilton Waikoloa Village, Hawaii, USA, 4–7 January 2017 (2017).


      9. Latos, B.A., Harlacher, M., Przybysz, P.M., M¨ utze-Niew¨ ohner, S.: Transformation

      of complexity drivers. In: 2017 IEEE International Conference on Industrial Engi- neering and Engineering Management (IEEM), pp. 1084–1088, December 2017.


      10. Schlagowski, R., Merkel, L., Meitinger, C.: Design of an assistant system for indus-

      trial maintenance tasks and implementation of a prototype using augmented real- ity. In: 2017 IEEE International Conference on Industrial Engineering and Engi- neering Management (IEEM), pp. 294–298, December 2017.



      11. Schuh, G., Riesener, M., Kantelberg, J., Steireif, N.: Transmission of software-

      related agile mechanisms of action towards product development processes for technical products. In: 2017 IEEE International Conference on Industrial Engi- neering and Engineering Management (IEEM), pp. 1753–1757, December 2017.


      12. Tams, S., Grover, V., Thatcher, J., Ahuja, M.: When modern technologies meet

      ageing workforces: older workers are more affected by demands from mobile inter- ruptions than their younger counterparts. In: 50th Hawaii International Confer- ence on System Sciences, HICSS 2017, Hilton Waikoloa Village, Hawaii, USA, 4–7 January 2017 (2017).


      13. Zhang, H., Liu, Q., Chen, X., Zhang, D., Leng, J.: A digital twin-based approach for

      designing and multi-objective optimization of hollow glass production line. IEEE

      Access 5, 26901–26911 (2017).


      14. Zheng, L., Bae, K.G.: An analytical study on horizontally collaborative transporta-

      tion strategies. In: 2017 IEEE International Conference on Industrial Engineering and Engineering Management (IEEM), pp. 564–567, December 2017.


      Smart Mirror Devices

      For Smart Home and Business

        Sven Von Hollen and Benjamin Reeh (&)


      Ernst-Abbe-Hochschule – University of Applied Sciences, 07745 Jena, Germany,

      Abstract. Smart home devices are more and more common these days. Not only

      normal computers are helping us to manage the everyday business. It is the way

      of interaction between computer and user, which is creating new value. To

      simplify the communication, the end user needs smart devices, such as smart-

      phones, smart door openers, smart washing machines or smart mirrors. The latter

      is investigated deeper in the following paragraphs. The study was performed in

      an university course – a technical student project, where the students not only

      created a concept, but also built a technical prototype of a smart mirror. During

      this process, the following tasks were accomplished: 1. Building the prototype/

      connecting all hardware components, 2. Programming a Raspberry-Pi computer

      in Python for the smart mirror device, including a camera with face scanning

      feature, different apps like weather forecast, live connected notebook/to-do list,

      welcome screen with personification and a clock, 3. Creating a short concept,

      where a market as well as economy analysis, presentation of the technical

      structure and analysis of the current application area was made. The result is an

      intelligent smart home mirror with personalized user interface. Keywords: Smart-home devices Smart mirror Prototype Intelligent devices

      1 Introduction

        Communication and especially the speed of communication between humans has increased enormously within the last decade. Today, there are millions of applications and devices for helping people to achieve their daily goals. Devices are defining our daily business and our daily life. Smartphones, for example, changed the way we have been living ever since and started a completely new communication era. Now we can book, plan, exchange, buy, sell and communicate everything within just one device, which has the size of a human hand. But not only the purpose of mobile phones changed, but also classical, normal objects in our daily life transformed into digital devices. One example is a typical bathroom mirror, which now can be transformed into an intelligent computer, a smart home device. This device can provide you with dif- ferent information, such as daily news, or personal notes from your smartphone. Within

        Smart Mirror Devices 195

        the project, which was a student project at the EAH – university of applied sciences, the motivation of the team was it to create value for users in the daily life at home. Today there is a bunch of information all over the web, which makes it complicated to get the right information in the right place at the right moment. The smart mirror can achieve this goal. Located in the bathroom, it can present information like daily tasks in the morning or weather forecasts in the evening. The goal was it to create such a prototype, which can be personalized via face recognition and which has different apps for the user interface. Additionally, there was a study about the current market situation, the fields where the device could be placed, the cost analysis of one prototype and already

      2 Related Work

        Smart home describes the equipment of residences with a technology that allows monitoring, management and maintenance of good health. They depend strongly on the situation of each person and vary by age, health and demography.

        As part of the smart home market, it is important to look at the status and devel- opment to see the current and future interest in smart mirrors. Looking at Europe, Germany and the UK have the biggest smart home market with around € 2,500 million each [

        The most popular applications

        are the ones which save time, such as control centers, automatic lighting/heating control or robot vacuum cleaners. Other uses of smart home devices include other ways of receiving/consuming information and safety. While for smart mirrors are also another, easy way to receive quick information, one of the main factors is the possi- bility to save time as you receive information on the fly (e.g. while brushing your teeth).

        Looking at the smart home market, there are four main services they are providing: health, security, energy and leisure (see Fig.

        As for healthcare, the main focus is the

        optimization of homes for elderly and disabled people. In the USA, the Georgia Institute developed a smart home which recognizes a potential crisis, the University of Florida developed a Gator Tech Smart House focusing on energy efficiency, safety, security, Reminder technologies as well as monitoring and in the UK developed an interactive dwelling house where a sensor system assesses vital signs and activities. It also provides security measures by controlling doors, windows and curtains


        Other facilities focus more on general optimization of homes, such as the PRO- SAFE project in France. In addition, there are many more countries and facilities which focus on the development and improvement of smart homes.

        Most of these researches focus on smart homes in general. The smart mirror cur- rently plays a minor part in smart homes as it has not penetrated the market yet. But looking at the technology behind each component there is a big interest in further development. Recognition of voices, faces and emotions play an important role in current and future technologies and will help improving smart mirrors over time.

        196 S. Von Hollen and B. Reeh Fig. 1. Most popular smart home applications in Germany 2017


      Fig. 2. Services provided by smart homes [


        Voice-Recognition While in the past voice recognition played only a minor role for smartphone users, the market introduction of home assistants such as Google Home, Amazon Echo or Apples HomePod changed the importance of voice interaction. New developments such as Viv dubbed as “The Global Brain”, the developers Cheyer and Kittlaus hope to enable ordering a pizza just by voice, without installing any other apps [

        In addition, the

        pure improvement of recognition involves a enormous improvement for future users, helping them to reach their daily goals.

        Smart Mirror Devices 197

        Face-Recognition After past security fails, especially due to terrorist attacks, countries started to research face-recognition as well as other safety mechanisms such as password-based authen- tication, iris scanning, and fingerprints had dangerous flaws. Face recognition repre- sented a good compromise between reliability and social acceptance

         ]. After initial

        research, there are still efforts to develop a more user-friendly system. Some factors

        The aging process causes many

        cause difficult challenges for face-recognition [ changes in facial textures and shape. Especially the slow aging process makes it challenging to train algorithms in this area. Facial expressions make scanning difficult the face. Regarding smart mirrors, this should be a minor problem as users are mostly focused on the mirror, yet other actions (e.g. brushing teeth) might cause similar difficulties. Especially problematic will be the variation of pose. People passing the mirror or taking other actions cause a lot of movement. With further improvement, smart mirrors should be able to scan faces in a variety of situations very fast to reduce frustration of users. Emotion-Recognition Advances in this field have found a significant place in communication between humans and computers. The main objective is for the computer to react accordingly to information by interpreted emotions [

        Researchers have incorporated facial

        expressions in devices to enable them to understand mental states. Many approaches for facial expression detection and recognition have already been proposed, training them with a variety of facial gestures. There have been many approaches focusing on facial expressions (appearance, shape, 3D tracking, poses), bodily expressions (body quantity of motion, face & body-displacement features, face & body segmentation), speech signals (acoustic features, gestures & speech, facial features), psychological signals (heart rate, facial features, biochemical features of breath gas, skin temperature) and other modalities (pupil diameter signals, 2D/3D features, pupillary response).

        Recognizing emotions plays a subordinate role for smart mirrors as the main usage is focused information. Yet in some situations, it might be useful that the mirror is able to react to the needs of users without direct orders. That way, the mirror can help with diseases or react accordingly to mood swings.

        Risks While there are many positive aspects of smart homes, there are also risks to look at. New technologies might cause additional problems or people could have difficulties using or accepting these devices.

        1. Initial contact with new technology [


        The first big obstacle to overcome is the initial contact with customers who know only little about smart homes. While there are many different devices, most customers want to solve a single problem, like automatic light regulation. The higher costs and difficulties of connecting these devices are too much of a hassle for many customers, so they avoid a purchase.

        2. Coexisting devices [

        As there are many different companies producing smart

        devices, at the moment these are using different protocols that might cause problems with communication between them. These communication issues are so complex,

        198 S. Von Hollen and B. Reeh

        3. Data security: Especially in Germany, people focus a lot on how safely data is used.

        Looking at Germany, when using cameras on private property, §27 of the Federal Data Protection Act excludes recordings for personal or family activities, so fam- ilies using cameras have no regulations using these. Meanwhile companies have to keep this act in mind. In addition to this comes the problem that some users might be scared when using these technologies, as they scan and film the person and their behavior. They might be suspicious of the company behind it and with that avoid a purchase. For companies it is important to communicate clearly with their cus- tomers to avoid any misunderstandings and assuring their data safety. This includes internet. Hackers could get important information about your life to prepare for example a robbery or collect personal data of yours.

        Customer Segmentation When approaching the market, there are big differences in interest and needs for each person. A survey looking at different marketing materials of the smart home market


        found that the most common characteristics looked at are the age and stage of the household. Most materials showed a white young to middle-aged couple with one child. In some cases, elderly where targeted. The decision-making for a purchase is very rational, there is a big focus on improvements such as minimizing energy use or maximizing convenience. Some marketing materials used a more emotional approach, especially for lights adapting to your mood and technology that helps elderly or dis- abled people (Fig.



      Fig. 3. Relative segment sizes for smart homes.

        Strategy Analytics published in 2017 a more specific segmentation into six groups [ (see Fig.

        Strugglers and Laggards are the biggest groups and both challenging to gain as they are not interested in new technology and low income.

        The other four segments can be achieved in different ways and are all interesting potential customers. Affluent Nesters are both: interested in new technologies and willing to invest in their homes. Impressers focus on newest technology to gain envy,

        Smart Mirror Devices 199

      Fig. 4. Components of the smart mirror in detail incl. all dimensions, (Translation:

      Spiegel = mirror, Rückwand = rear panel). (Color figure online)

        they focus on products that leave a big impression and are rare to see in current households. They are also a core segment for the smart mirror market as these might surprise visitors passing them. Early adopters are not scared to be the first approaching new technology while it is still unknown to most of the population. Also, they are willing to pay higher prices. This segment has a very small percentage but a high priority as they begin to make the products more commonly known in public and giving valuable feedback for further product improvements. Greens are a bit different to approach, as their focus is on environmental improvements. They focus on technology that will save energy or help the environment in any other way. While they are not early adapters, as soon as the product penetrates the market, they are willing to invest some money.

      3 Approach

        In the following section, the prototype of the smart mirror, including all components, are presented. Moreover, the total development process like the setup of the smart mirror hardware is investigated. In the following illustration, the main parts of the smart mirror are listed:

        On the left side there is the rear panel, which is holding all components together, just like a casing. The casing of course always depends on the end user and where the device is integrated. So, the casing would be different for example in a bathroom or toilet. In the middle, there is the display, which was taken from a computer screen. Below the display, there is the display controller (green), which provides the display

        200 S. Von Hollen and B. Reeh

        connection for the Raspberry Pi model 3 – the main computer (red). For simple use cases, a Raspberry Pi is working without any problems. For more complicated use cases, the Pi is to slow, and the performance is not adequate. Finally, on the right side there is the mirror, which is responsible for the modern design of the prototype. Because there was not enough budget in the student project, the mirror was built out of a Plexiglas, combined with a mirror foil. For a good standard it would be better to use high quality mirror glass/spyglass. The mirror was connected with glue to the display, the computer was attached to the back of the screen. Also because of financial restrictions, there was no frame/no back panel used in the project (Fig.



      Fig. 5. Building the smart mirror and connecting the hardware.

        The software was running on a raspberry pi model 3 with an ubuntu system and used like a webserver with a html frontend. Different “Apps” like clock, connected to Do List and weather forecast are connected to the internet and will update every 2–3 s. Different functions where implemented with iframes. Moreover, within the project, different tools with face recognition were tested. One of them is the mood recognition tool. The software can use the camera to understand the user’s feelings. For every mood the software shows a ranking, how accurate the software measures the user’s mood. For measuring there are different categories for moods: WUT = fury,

        VERACHTUNG = contempt, EKEL = disgust, ANGST = fear, FREUDE = joy, NEUTRAL = neutral, TRAUER = grief, ÜBERRASCHUNG = astonishment; The mood recognition software was just tested, but not implemented in the system because of project time restrictions. The face recognition iFrame was implemented successfully (Figs.


        Smart Mirror Devices 201

      Fig. 6. Mood recognition via camera face recognition iframe.


      Fig. 7. Smart mirror prototype, first hardware version of the smart mirror.

      4 Discussion and Future Work

        Obviously, the smart mirror device can be used in many different areas. As already mentioned, it could be a bathroom device or used in a toilet. Additionally, there are many other commercial fields for the device, which will be presented in the following paragraph. Because of new device areas, there will be discussions about the future work for new prototypes in different use cases with different conditions.

        With many different options of configuration (camera, microphone, apps, adding external devices which is very easy in general), the smart mirror holds a potential for many use cases. Firstly, smart mirrors can be used as a smart home device. For

        202 S. Von Hollen and B. Reeh

        example, in the floor, close to the front door. Normally every usual household has a mirror in the corridor, for checking yourself before leaving the house. Therefore, it would be a luxury to get a weather forecast or temperature details on the screen. In a connected smart home, this mirror could also show a map, with an overview of the household, for checking if all lights are switched off and if all windows are closed. A camera can be used for recognizing people and load their personalized apps and modules, for example to do lists (Fig.


        Fig. 8. Smart Mirror for a use case in the corridor [

        As already mentioned, there is also the option of a smart mirror located in the bathroom. Early in the morning while brushing your teeth, the smart mirror shows latest news, to dos and meetings for today or starts playing oneˋs favorite radio. Users could be identified via face recognition or without a camera module with connection to a smart phone. This can help users to start their day with all information and service they need (Fig.


        Fig. 9. Smart Mirror for a use case in the bathroom ].

        Another use case would be the smart mirror device for business use. For example, in a barber shop the mirror could be a helpful device. Via camera the mirror could project examples of haircuts or styling, so the customer could see their hairstyle before wearing it. Additionally, the customer could enjoy entertainment, such as short clips or reading the news on the smart device during his haircut. Samsung has already created such mirrors for barber shops ] (Fig.



      1. Statista. .

        Accessed 8 Mar 2018


      2. Statista.

      . Accessed 8 Mar 2018


      3. Chan, M., Campo, E., Estève, D., Fourniols, J.-Y.: Smart homes – current features and

      futures perspectives. Elsevier J., 91–92 (2009)


      4. Hargreaves, T., Wilson, C.: Who uses smart home technologies? Representation of users by

      the smart home industry. School Environ. Sci. J., 1773–1779 (2013)


      5. Toronto Sun.

      Accessed 8 Mar 2018

      6. Abate, A.F., Nappi, M., Riccio, D., Sabatino, G.: 2D and 3D face recognition: a survey.

        Elsevier J., 1885–1886 (2007)

      7. Sharif, M., Naz, F., Yasmin, M., Shahid, M.A., Rehman, A.: Face recognition: a survey.

      J. Eng. Sci. Technol. Rev., 166–168 (2017)


      8. Vinola, C., Vimaladevi, K.: A survey on human emotion recognition approaches. Databases

      Appl. J., 24–33 (2015)


      9. Theinstitute (IEEE).

      Accessed 13 Mar 2018

        Fig. 10. Smart Mirror Device for barber shops [

        Smart Mirror Devices 203

        204 S. Von Hollen and B. Reeh


      10. Strategy Analytics.

        Accessed 13 Mar 2018


      11. Smart mirror in the bathroom. Acces-

      sed 16 Mar 2018


      12. Samsung barber mirror.

      . Accessed 15 Mar 2018


      13. Smart mirror in the corridor.

      . Accessed 18 Mar 2018



        . Accessed 18 Mar 2018



      Accessed 18 Mar 2018


      16. Accessed 18 Mar


        Short Papers: Security and Systems


      Secured Domain of Sensor

      Nodes - A New Concept

        Janusz Furtak (& , Zbigniew Zieliń

        , and Jan Chudzikiew


      Military University of Technology, Warsaw, Poland



      Abstract. For the procedures of creating situational awareness an important

      element is the freshness (timeliness) of data and the credibility of information

      sources. A modern and rich source of such data may be sensor networks, but

      their use as a reliable source of data is a major challenge mainly because the

      sensor nodes in the network are usually mobile, use wireless links, have a small

      computing power and have small energy resources. The paper presents a new

      concept of a secure sensor nodes domain, which is a very reliable source of

      sensor data. The data transferred inside the domain are protected by cryptog-

      raphy, cryptographic material stored in the resources of sensor nodes is also

      protected by cryptography, and data obtained from sensor nodes from the place

      of their acquisition to the sink node are protected. Trusted Platform Module

      (TPM) is used to support the process of securing the data in the domain. In the

      second part of the paper developed a secured domain demonstrator for sensor

      nodes is described, which illustrates the functioning of the key elements of the

      concept presented.

        Keywords: Wireless sensor networks Security in IoT Trusted Platform Module

      1 Introduction

        Observing today’s development of Internet technologies, there is no doubt that in the near future the number of network nodes that will be the source of data and the number of recipients of this data will grow rapidly. If the recipient of the data will be able to use this large amount of data, and especially the data that is the most fresh and up-to-date, it will get a new quality that has not been possible to obtain until now. Unfortunately, together with the advantages of broad access to very up-to-date data, threats appear associated with trust in this data. Using the Internet of Things, you can have easy access to publicly available data, but then you often lack information on the origin of this data, which in the case of critical infrastructure of the organizations or the state can be very important. Certainly, in the future, risk analysis methodologies will have to take into account data reliability depending on the source (e.g. in government appli- cations), data freshness understood as the moment of obtaining such data (e.g. in natural disaster management systems), or both in military applications.

        208 J. Furtak et al.

        This paper describes a new approach to building a network of sensor nodes , which is a very reliable source of data. The entire network creates a secure domain in which data transfer is protected by cryptography, cryptographic material stored in the resources of sensor nodes is also protected by cryptography, and data obtained from sensor nodes from the place of their acquisition to the sink node are protected. Trusted Platform Module (TPM) is used to support the process of securing the data in the domain [



        In the first of these solutions, the transmission security was based on asymmetric cryptography (RSA). Unfortunately, this approach required relatively large memory resources and high The second approach

         ] used asymmetric cryptography (RSA) only to build a trust

        structure in the domain, but symmetrical cryptography (AES) was used in the remaining operations. In those solution, there was one sensor node in the domain that was responsible for both preparing the domain’s nodes and the current operation of the domain. From a security point of view, this sensor node was a very sensitive element of the domain.

        In the presented approach, sensor nodes, which are to create a domain, are first prepared in a safe and controlled space outside their normal work area. It is only later, when they are moved to unprotected space, will start work there. All nodes are equipped with TPM modules. All transmission and sensitive domain data are protected by cryptographic methods supported by TPM. The equipment of sensor nodes allows to build mechanisms to detect unauthorized interference in the equipment of the node and its software.

        The rest of the paper is organized as follows. In Sect.

        the concept of

        secured domain is explained. Section describes an implementation, Sect.

         a short

        view on demo of secured domain and obtained results. Section

         presents some con- cluding remarks.

      2 Motivation

        It is very difficult to define what is meant by the phrase “secured network of sensors”, which is part of the Internet of Things. You can look at this problem from different points of view and take into account different aspects. Certainly it will be possible to say about the network that it is secured if it includes only secure sensor nodes that have their identifiers and that are authenticated before starting its work in the network, and the exchange of data between such nodes is secured. Such a network should also be resistant to failures understood as stopping the proper operation of individual nodes of this network, leaving the mobile network node out of range of the communication link 1 used, or the battery discharging, which supplies the sensor node.


      Sensor node - the element of sensor network, which includes at least the measuring element (sensor),

      microcontroller and communication module that allows transfer of measured data through wireless


        Secured Domain of Sensor Nodes - A New Concept 209

        The protected sensor node should be resistant to external and unauthorized impact on the sensor node consisting of modification of the sensor node’s equipment (e.g. replacement of sensor, memory elements or connection of additional equipment to monitor the operation of this node, etc.), change the sensor node’s software (e.g. to enforce operation other than the original one) or manipulation of data stored in the sensor node’s resources. The sensor node should be able to detect such activities and have the ability to respond to such actions, for example by sending a message to other sensor nodes about this fact, blocking its operation, destroying the data resources stored in its resources, or even physically destroying itself. To exchange data in a secured is not sent in explicit text. In addition, the protocols used in such a network should give the possibility to use mechanisms that make the network immune to attacks such as: Replay Attacks, Man-in-the-Middle, spoofing and others.

        Many of the described requirements can be met by using asymmetric cryptography (e.g. RSA or ECC) [


      ]. It is important to

        realize that using cryptographic techniques there is a problem of secure storage of cryptographic keys, the problem of secure distribution of these keys and the need to build a trust structure in a secured network. Such solutions are known and used in the Internet, e.g. a Certification Authority (CA).

      3 The Concept of Secured Domain

        Network of sensor nodes are usually mobile, uses a low-bandwidth and low-range communication medium, have relatively small memory resources, and low computing capabilities and limited power capabilities. For such nodes, it is not possible to effectively use, for example, the Certification Authority (CA) to build a trust structure. In this case, the mechanisms offered by the TPM module, which should be an obli- gatory component of each sensor node, can be used to build a trust structure.

        Due to small resources of sensor nodes, mobility of these nodes and small range of the transmission medium used, securing the sensor nodes network should be organized locally in small clusters of sensor nodes


        One cluster can represent one group of cooperating objects (sensor nodes). Such a object can be one soldier who is equipped with various sensors (e.g. body sensors). The data from these sensors are collected by the device (this device will be referred to as the sensor node) owned by the soldier. Each sensor node is mobile, for data exchange with other sensor nodes uses wireless connection and each sensor node is able to perform the same functions from the point of view of the sensor nodes network. These functions include:

      • gather and preliminary process of the data deriving from sensors located in the each of the objects;
      • perform protected transfer of data collected by objects of one cluster to other clusters;
      • perform a diagnostics and reconfiguration between objects of the cluster;
      • ensure transmission security within the cluster and authentication of objects in the

        210 J. Furtak et al.


      Fig. 1. Structure of one Sensor Nodes Network Cluster

        It was assumed that in each cluster there are two interpenetrating domains to which all cluster nodes belong:

      • Security Domain. In the domain exists exactly one sensor node (M node), which plays the Master role for the domain and is the authority in the domain. Other sensor nodes (R nodes) of domain exchange with M node the data used to authenticate the nodes in the domain. A symmetric key (NSK - Node Secure Key), which is used to secure transmission between a given node and the M node, is known only to these two nodes. For this reason, despite using the XBee link, the logical network topology is the topology of the star (Fig.

      2 A). In this domain are also implemented

        the diagnostic procedures. Then another symmetric key (NDK - Node Diagnostic Key) is used, which is common to all members of the domain.


      Fig. 2. Logical topology of network in security domain (A) and network in domain of

      transmission of the data obtained from sensor nodes (B)

      • Domain of Protected Transmission. The data obtained from all sensor nodes of the cluster are transferred to one node (G node), which plays the Gateway role in the domain. During this operation, a symmetric key (NTK Node Transfer Key) is used that is common to all members of the domain. The G node is also responsible for protected transfer of the data to other clusters. Logical topology of network in the

        Secured Domain of Sensor Nodes - A New Concept 211

        When the diagnostic procedures determine that the node that plays the role of Master or Gateway has failed or is not reliable, then the node selection procedure is started, which should take over the role of Master or Gateway respectively. It is acceptable that the role of Master and Gateway can be played by the same node, but it is not recommended.

      4 Implementation of Secured Domain

        In order to build a secure domain of sensor nodes that meets the postulates described in the previous section, the actions should be taken, which are described in the following subsections. It was also assumed that each domain node will be equipped with a TPM and a local root of trust will be created on each node with support of TPM. The local root of trust will be based on the key hierarchy, at the top of which is Endorsement Key (EK). The next key in the hierarchy is the Storage Root Key (SRK), whose descendant is Domain Key (DK). All keys forming a root of trust are RSA 2048 asymmetric keys. The EK is created once and it is not possible to change nor delete it from the resources of the given TPM.

        Its private part is in no way accessible to the surroundings of the TPM. The SRK is generated during the takeover procedure for ownership of the TPM module. During this procedure, a secret sequence is used, the knowledge of which will enable performing authorized activities with a given sensor node. The private part of this key is protected by the EK and it never leaves the TPM. This key can be regenerated, but then all previously used cryptographic material is lost. The DK is the key shared by all sensor nodes of the secured domain. In a special procedure, it will be securely transferred to other domain nodes.

        Taking into account the above observations, the actions related to the creation and use of a secured sensor nodes network can be divided into the following organizational activities:

      • activities related to: the preparation and use of network sensor nodes, protection of the software, the equipment of node and data there stored, • activities related to securing transmission links used in the sensor nodes network.

        4.1 Architecture of the Sensor Node It has been assumed that the sensor node (Fig.

         A) will be a mobile device powered

        from its own power source and equipped with an Arduino microcontroller, TPM 2 module and LoRA interface connected via I C bus, XBee interface connected via Serial link and one or more sensors that will be source of data. The LoRa interface is designed to communicate with the cluster’s external environment in the event that a given node 2 plays a Gateway role. The XBee interface is designed to exchange data inside the


      This implementation of a secure domain demonstrator uses the TPM 1.2, which offers support for

      RSA 2048, but there is nothing to prevent the use of the TPM 2.0, which allows the use of ECC

        212 J. Furtak et al.

        domain. The Serial2 port is used only in the procedure of preparing a node to work in the domain. During this procedure, the node being prepared will be directly connected to node B node, the diagram of which is shown in the Fig.



      Fig. 3. Block diagram of R node and B node

        4.2 Organizational Activities The main motivation to create a secured domain of sensors is the fact that this domain will operate in unfriendly environment for network of sensors. Undoubtedly the bat- tlefield is a such environment. Therefore, every element of such a domain must be early prepared to work in such an environment before it goes into this hostile environment.

        Therefore, it was assumed that the preparatory activities should be carried out in a safe environment outside the area of normal network operation. The Base node (B node) is used to perform the preparatory activities.

        Preparatory activities include the following steps: 1. Preparation of the node (B node) and generation of data for the security domain.

        2. Preparing a node that will act as a Master in the security domain.

        3. Preparation of the remaining nodes that will act as a Replica in the security domain.

        The first step of creating a secure domain of sensors is the initialization data generation for the domain sensors. The structure of this data and the place of its storage in the resources of B node are shown in Fig.


        The data stored in the NVRAM of the TPM module are protected by the mecha- nisms of this module. These data include:

      • N_ID - B node identifier
      • NK - symmetric key to encrypt the data stored in the RAM and EEPROM of the node
      • PAN_ID, CH, PIK - transmission parameters for XBee interfaces that will be used in the domain.

        The data stored in the EEPROM memory are protected with the NK key of the node and contains a description of sensors’ domain (i.e. domain name, current version of the domain description and time parameters related to domain copy creation) and generated identifiers for nodes of the domain. These identifiers will be forwarded to these nodes during the procedure for preparing these nodes. After completing the node preparation

        Secured Domain of Sensor Nodes - A New Concept 213

      TPM Non-volatile Memory

      DK SRK EK





        DK Key_Desc N_ID NAD DN RN

      EEPROM of node RAM of node

      VER PR PNR TDV OTP domain description N_ID RN IV N_ID NAD

      Fig. 4. The data stored on B node

        procedure, the XBee interface address that the given node will use, will be entered into the appropriate NAD field.

        In the next step, all sensor nodes should be prepared for work in the domain. During the node preparation procedure, the local trust structure (EK, SRK) is generated first, and then the NK key for the node. In the next step, the DK key is migrated in a secure way using the Serial2 connection from the B node to the node being prepared.

        Next, using the public part of the transferred DK key from B node, the following data are acquired: transmission parameters (PAN_ID, CH and PIK) for the XBee interface, the address of the node acting as the Master in the domain and the individual node tag (NTAG). This tag will be verified in a later node registration procedure in the domain.

        It was assumed that the first node being prepared would be a M node in the domain. For this reason, during the procedure of preparing this node, additionally a description of the domain is sent to its resources. The contents of the M node and R node resources are shown respectively in the Fig.


        4.3 Procedures in the Security Domain Only the node that previously successfully completed the registration procedure in the domain can function in the security domain. The condition of proper registration in the domain is earlier preparation of this node by node B node (i.e. saving in the resources of this node the credentials necessary for registration in a given domain).

        During normal operation of the sensor nodes domain, only wireless links are used: the XBee link to exchange data inside the domain and the LoRa link to transfer data outside the domain through the node acting as the Gateway. Data contents are encrypted during data transmission. In the case of data transmission between M node and other nodes, individual NSK keys of these nodes are used. During the transmission of data obtained from sensor to the G node, the NTK key is used, which is common for the entire domain. During the diagnostic procedures an NDK key is used, which is also common for the whole domain. The content of these keys (NSK, NTK and NDK) and the G node address (GAD) are determined during the node registration procedure in the The following procedures are necessary for the functioning of a secure sensor nodes domain:

        1. The procedure for initiating M node - creation of data structures in the node resources, which is to act as a Master in the domain.

        2. The procedure of restoring M node and the entire description of the security domain - recovery and verification of data after power on M node.

        3. The procedure for registration a node in the security domain.

        4. The procedure for node status confirmation - the procedure starts when the node is powered up.

        5. The procedure for sending data from R node to G node.

        6. The procedure for sending data from G node outside the domain.

        7. Diagnostic procedure for domain.

        8. The procedure for electing of new M node.

        9. The procedure for deleting nodes of domain on M node.

        10. The procedure for electing of new G node.

        4.4 Demo of the Security Domain The domain demonstrator, which includes four sensor nodes, was prepared to verify the proposed concept of a secured domain of sensors. The presented demonstrator has implemented all the preparation procedures for sensor nodes of domain, which are described in Subsect.

        and the first six procedures described in Subsect.


        Each of four nodes used in demonstrator will be a source of sensor data that will be passed outside the domain. Among these nodes, one will be the MASTER and one will be the GATEWAY. TPM Non-volatile Memory RAM of node NAD MAD GAD PIK PAN_ID CH EK SRK DK EEPROM of node Key_Desc DN PR PNR TDV ”M” VER domain description N_ID RN NAD NSK IV N_ID RN NAD NSK IV N_ID RN NAD NSK IV node description N_ID MAD NAD RN SQ GAD N_ID NTAG NDK IV NK IV NSK IV NTK IV TPM Non-volatile Memory RAM of node EEPROM of node Key_Desc NAD MAD GAD PIK PAN_ID CH EK SRK DK N_ID MAD NAD RN SQ GAD N_ID NTAG NDK IV NK IV NSK IV NTK IV



      Fig. 5. The data stored on M node (A) and R node (B) after preparing procedure 214 J. Furtak et al. Secured Domain of Sensor Nodes - A New Concept 215

        Each sensor node (Fig.

      6 A) used in the experiments was built with the following

        components: microcontroller Arduino Mega 2560 R3, Cryptotronix CryptoShield containing TPM v.1.2, Dragino LoRa Shield - 868 MHz v1.3, wireless communica- tion module XBee, and ultrasonic distance sensor. In the procedure of preparing sensor nodes to work in the domain, the BASE node Fig.

         B) is necessary. It was built with

        the following components: microcontroller Arduino Mega 2560 R3 and Cryptotronix CryptoShield with TPM v.1.2. Figure

      6 C shows the configuration of the demonstrator during the procedure of preparing the sensor node for work in the domain.


      Fig. 6. View of sensor node (A), BASE node (B), and sensor node during the procedure of

      preparing the node for work in the domain.

        For the purpose of experiment, a domain (Fig.

        included four sensor nodes was

        prepared: one of then served as Master, and the other served as Gateway. Several test have been carried out to check to what extent encryption of the content of transmitted data affects the delay of handling individual frames by G node. The time needed to handle encrypted frames is about 83 ms longer then handling unencrypted frames – see the result obtained in Table 3 Fig. 7. Configuration of the demonstrator during the transmission of sensor data to G node

      Used module meets the requirements described in Security Policy for Atmel TPM 9, as well as the

        216 J. Furtak et al.


      Table 1. Transfer time of sensor data acquired by Gateway node

      SD Protect Length [b] sensor_packet_req Transfer sensor_packet_ans [mm] [ms] Support Difference Prepare Difference

        [ms] [ms] [ms] [ms]

        56 NO 22 7,13 80,16 12,91 0,28 86,59 AES

      22 87,29 12,94 86,87

        5 Conclusion

        The presented demonstrator configuration presents the basic functions of the secure sensor nodes domain. The equipment of each sensor node allows to build a network including not more than sixty one sensor nodes. Sensor data obtained from domain nodes in a secure manner can be passed inside the domain and to the recipient from outside the domain. The presented domain is one of the elements of the Application of IoT in Military Operations in a Smart City, which is described in

         ]. The next step in

        improving the domain is equipping it with a Quantis RNG hardware random number generator

         ] that provides full entropy.

        Until now, not all required domain functions have been implemented. These functions include: diagnostic procedures that should detect damaged sensor nodes or those nodes that have lost confidence, and the procedures for the election of new M node or G node in the case of detected disability of node that have performed this role until now. These tasks and securing each node against unauthorized interference will be the subject of the forthcoming work.



      1. Al-Haija, Q.A., Tarayrah, M.A., Al-Qadeeb, H., Al-Lwaimi, A.: Tiny RSA cryptosystem

      based on arduino microcontroller useful for small scale networks. Procedia Comput. Sci. 34,

      639–646 (2014).


      2. Furtak, J., Zieliński, Z., Chudzikiewicz, J.: Secure protocol for wireless communication

      within military IoT. In: 2015 IEEE 2nd World Forum on Internet of Things, WF-IoT 2015,

      pp. 508–513 (2015)


      3. Furtak, J., Zieliński, Z., Chudzikiewicz, J.: Security techniques for the WSN link layer within

      military IoT. In: 2016 IEEE 3rd World Forum on Internet of Things, WF-IoT 2016, pp. 233–

      238 (2016)


      4. IDQ, Quantis AIS31 Brochure


      5. Jonsen, F.T., et al.: Application of IoT in Military Operations in a Smart City presented at the


      IEEE International Conference on 2018 Military Communications and Information Systems



      6. Koteshwara, S., Das, A.: Comparative study of authenticated encryption targeting lightweight

        IoT applications. IEEE Des. Tests 34(4), 26–33 (2017).



      7. Requirements For Cryptographic Modules. Federal Information Processing Standard (FIPS

      140-2), National Institute of Standards and Technology, 03 December 2002

      8. Puschmann, D., Barnaghi, P., Tafazolli, R.: Adaptive clustering for dynamic IoT data streams.

        IEEE Internet of Things J. 4(1), 64–74 (2017)

      9. Trusted Computing Group. TPM Main Part 1 Design Principles. Specification Version 1.2.

      Revision 116, Trusted Computing Group, Incorporated (2011) Secured Domain of Sensor Nodes - A New Concept 217


      Trends in Application of Machine Learning

      to Network-Based Intrusion Detection Systems


        Jakub Hrabovsky , Pavel Segec , Marek Morav


      University of Zilina, 010 26 Zilina, Slovakia



      Abstract. Computer networks play an important role in modern industrial

      environments, as many of their areas heavily depend on continued operation and

      availability of provided network services. However, the network itself faces

      many security challenges in the form of various massive attacks that prevent its

      usage and yearly cause huge financial losses. The most widespread examples of

      such devastating attacks are the Denial of Service (DoS) and Distributed DoS

      attacks (DDoS). This paper is focusing on the analysis of detection methods that

      eliminate attacks impact. The paper introduces challenges of the current network

      based intrusion detection systems (NIDS) from distinct perspectives. Its primary

      focus is on the general functionality of selected detection methods, their cate-

      gorization and following proposal of some potential improvements. Considering

      the requirements on present and future NIDS, we emphasize the application of

      machine learning (ML). The paper analyzes the state of research of four par-

      ticular ML techniques regarding their success in implementation as NIDS –

      Bayesian Networks (BN), Support Vector Machines (SVM), Artificial Neural

      Networks (ANN) and Self-organizing Maps (SOM). The analysis reveals vari-

      ous drawbacks and benefits of the individual methods. Its purpose lies in the

      discovery of current trends showing a direction of the future research, which

      may possibly lead to the overall design improvement of new methods. The

      output of our research summarizes trends in the form of trends list and their

      influence on our future research.

        Keywords: DoS DDoS Intrusion detection NIDS Anomaly-based Machine learning

      1 Introduction

        The increasing role of the Information and Communication Technologies (ICT) within new areas of industry, such as Cloud Computing (CC) or Internet of Things (IoT), makes computer networks the most important part of their core infrastructures. There- fore, computer networks have to meet the requirement of full availability because of the close dependency on them. Their potential unavailability causes financial losses and is a threat in the critical deployment areas like health care and the power engineering.

        The availability is one of three primary properties of secure network (beside the con- fidentiality and integrity) and therefore its maintenance is a problem of computer security.

        Nowadays, from the point of distinct views, the computer network security deals with the following problematic areas and related issues [


      • Implementation view: The development of bug-free software systems is an impossible challenge because of the requirements and costs it would require. It means, that the software has always unknown security holes and vulnerabilities that can be misused by some attacker, now or in the future.
      • Attacker’s view: The attackers improve their skills and develop new attacks the same way and at the same rate as the new detection methods arise. Therefore, the threat space is very huge according to the openness and variability of the network and its services.
      • Victim’s view: The TCP/IP architecture and a related global network, the Internet, are based on the principle of openness that forces the devices to process all incoming traffic regardless of its origin. It also enables a disguise of the original identity through the crafting and sending of packets with custom content. That makes the identification and tracking of an attacker much more difficult.
      • Completeness view: Despite a number of existing methods, no universal method exists that provides a complete security and defense protection of the network against all possible threats.
      • Network infrastructure view: The current network infrastructure is a complex and very dynamic system. The environment is changing rapidly with the arrival of new technologies that increase the number of connected devices, link bandwidth, and amount of transferred data. These changes place higher demands on methods and their performance than it was in the past.

        Current methods do not address issues mentioned above and so there is a need to develop novel methods of prevention, detection and response to the network attacks to keep up with attackers.

        This paper analyses several approaches to previously described issues mainly based on detection techniques. The aim of the paper is to find trends in current methods through the comparison of their properties and applied concepts. The second section of the paper provides a general background of the network intrusion detection problematic including definition of Denial of Service (DoS) and Distributed DoS (DDoS) attacks and hierarchical classification of common intrusion detection methods. The third section gives the relevant research overview of Bayesian Networks (BN), Support Vector Machines (SVM), Artificial Neural Networks (ANN) and Self-organizing Maps (SOM) application as a part of network intrusion detection system. The section also presents current challenges in the method performance evaluation. The results of the paper are summarized as the list of trends in the discussion section.

        Trends in Application of Machine Learning 219

        220 J. Hrabovsky et al.

      2 Background of Network Attacks and Their Detection

        DoS and DDoS attacks are well-known network attacks that are able to partially or fully block network services [

         ]. Despite the intensive research in the area of

        network security, impact of these attacks increases each year. Many official reports

        The primary goal of the DoS/DDoS attack is the direct or

        confirm this trend [ indirect exhaustion of network (bandwidth), memory (HDD, RAM) or computing (CPU) resources on the victim side through the deliberate activities of attacker. The affected service is partially or fully unavailable, which greatly decreases its quality. The victim can be an end device (client, server), an intermediary (router, switch) or a network channel (link).

        The current research and solutions to preceding network attacks suggest four approaches: detection, prevention, mitigation and response

         ]. The early detection can

        decrease the effect of an attack on the service quality because it leads to the immediate reaction. The principle of intrusion detection methods lies in comparison of gathered data from a monitored environment with a model. The model serves as a familiar pattern in the process of differentiation. The final method decision depends on the evaluated similarity between the current data and the pattern, and on the behavior that the model represents.

        The taxonomy of intrusion detection methods varies in parameters like type of processed data and place of deployment



        2.1 Intrusion Detection Systems Classification Based on Data Sources From the perspective of data sources and the placement, two common method classes are known: host-based and network-based intrusion detection systems (IDS) [


        Host-Based IDS (HIDS) Host-based IDS (HIDS)

         ] is deployed at the end host. HIDS uses data gathered by

        operating system and services running in end devices (client PCs, servers, printers etc.) or in intermediate nodes (switches, routers and firewalls). Data are stored in the form of logs

         ]. System under control has to be monitored continuously, so potential threats

        can be detected early enough. Host-based methods analyze gathered logs and create overview of system state [

        This type of methods focuses on the detection of attack

        types that impact only the system itself: Remote-to-Local (R2L) and User-to-Root (U2R) attacks, in general. Anti-virus systems, firewalls and anti-spyware programs are the examples of HIDS.

        Network-Based IDS (NIDS) Network-based IDS (NIDS) uses data gathered from the surrounding network envi- ronment in the form of captured packets. Primary role of NIDS is to distinguish attack from normal traffic and to identify packets or flows belonging to the attack. Main task of NIDS lies in creation of a model that represents a particular kind of network traffic behavior - its pro

        file. The full process of detection has to be done early enough, as the potential attack affects service quality (bandwidth of network connection, response latency of network services, etc.).

        Trends in Application of Machine Learning 221

        The second mentioned operation, the identification, is equally important and nec- essary to perform subsequent countermeasures, like filtering and traffic limiting. When the method detects an attack, it has to find out suspicious part of the traffic and specify its parameters.

        Dominant type of attacks that are detectable by NIDS are DoS/DDoS and various forms of reconnaissance (scanning of network and searching for information about connected devices).

        2.2 Intrusion Detection Systems Classification Based on the Analysis Approach

        From the perspective of model representation, IDS methods can be classified into two groups: misuse-based (aka knowledge or signature-based) and anomaly-based (aka behavior-based)

         ]. A model created by misuse-based method stands for suspicious

        behavior. On the other side, anomaly-based methods perceive model as a sufficient representative of normal behavior. Categories in the following sections are described from the NIDS perspective. Misuse-Based Methods Misuse-based methods are highly dependent on the knowledge of existing attacks saved in database. The database consists of profiles of all well-known attacks described by signatures. The signature is an ordered list of parameters or features with their specific values and ultimately distinguishes the corresponding attack from others. For example, in the case of R2L and U2R attacks it is possible to use the particular order of system calls as the unique feature. Another example are values of the IP packet header fields captured from network traffic.

        The key idea of misuse-based detection lies in the analysis of samples gathered during traffic monitoring. The method compares samples with signature rules loaded from database. When the equality occurs, the sample is tagged as the attack corre- sponding to the positive signature. Because the signatures describe attacks in detail, the accuracy of detection and classification is very high. To achieve such a high precision, the database may consist of a huge number of entries (i.e. attack signatures). The representation, processing and evaluation of database entries differs from method to method.

        The method performance depends on signature database and requires its regular updates. Updates are performed by security experts, who are responsible for adding signatures of new attack types. Consequently, the quality of database content and achieved results depend highly on the knowledge and experience of security experts. This can be considered as the serious disadvantage and reason of moving to another detection class - anomaly-based detection.

        Anomaly-Based Methods Anomaly-based methods are a counterpart of misuse-based methods because they create a model of network without any suspicious traffic. The model generation process utilizes the activities of devices in the monitored network in order to create the exact representation of the environment in its normal state. Such model has to be updated

        222 J. Hrabovsky et al.

        regularly to stay up-to-date because the behavior of current network infrastructure is very dynamic.

        Methods identify the anomaly as the significant difference from the normal network activity specified by the model. The malfunctioned devices and flooding network attacks are examples of network anomalies. The model form and technique of its creation depend on the chosen method.

        In order to get valuable results, the method supposes that model performs dynamic updates. To do that, there actually exist various sets of network traffic samples, called datasets, and algorithms applicable for processing of big data that support the process training). The quality of model training depends on the learning algorithm and on the size of a dataset.

        Datasets also play significant role in testing, where the quality of a model is mea- sured. In general, a bigger dataset means a more accurate trained model regardless of used algorithm. Samples in a dataset can be labeled as a normal traffic or a particular type of anomaly. To make labeling correctly is a difficult process that requires time and experience of security experts. We differentiate three modes of learning based on the proportion of labeled to unlabeled samples in dataset:

      • The supervised learning uses a dataset consisting of only labeled samples. The model has then all known classes at the disposal and thus enables the full classification.
      • Datasets used in semi-supervised learning split samples into two groups: labeled and unlabeled. As a consequence, the accurate multi-class classification is more difficult.
      • The unsupervised learning uses only datasets with unlabeled samples and lacks additional information (no labels). Therefore, the creation of a model is more dif- ficult compared with previous learning modes. On the other side, this mode enables use of real-time network traffic to learn the model with good accuracy.

        The ability to detect beside of known attacks also their modifications or totally new attacks (zero-day attack) is important advantage of anomaly-based methods nowadays. On the other side, the accuracy of anomaly-based method is lower compared with misuse-based methods because it finds anomalies instead of direct attacks. Assuming attacks only as a subset of anomalies, we often incorrectly evaluate any dynamic change of user or device behavior as an attack. Consequently, anomaly-based methods suffer from frequent false positives.

        Anomaly-based methods offer a space for many adjustable options and learning process that have great impact on the results. From this point of view, two main approaches are common: statistical analysis and machine learning (ML). According to the paper focus, only machine learning is addressed in the following sections.

        Machine Learning Present computer network systems are too complex in order to model them exactly. ML methods allow us to create an approximate model from input samples only, without any knowledge of the internal system behavior. They need to select hyper-parameters that specify overall structure of a model but its concrete behavior is forming automatically

        Trends in Application of Machine Learning 223

        through learning. The model learns new patterns from input samples in order to identify similar or modified inputs in the future. This property, called generalization, is very important advantage of ML, especially in detection of anomalies. Iterative process of continuous learning improves the quality of a model and its results. Every ML method is defined through its model, parameters and an error function. The training does parameter adjustment with the goal to decrease the final error of every sample evaluated by an error function. Two main method categories are known in the field of ML: classification/regression and clustering.

        Some of ML methods proved their successful implementation in the area of net- work anomaly detection – Bayesian Networks [


        Artificial Neural Networks, and Self-Organizing Maps

        3 Machine Learning Methods in NIDS and Their Evaluation Challenges

        This section provides overview of some research papers supporting the application of above mentioned machine learning methods in NIDS domain. Subsequently, various evaluation challenges are presented that misrepresent the performance and comparison of existing methods.

        3.1 ML Methods in NIDS Bayesian Network (BN) and its lightweight derivation – Naive Bayes (NB) – are

        The authors prioritize a hybrid method con-

        sisting of multiple simple, but specialized NB models. Such ensemble models enable to connect prior knowledge with training process and thus to improve the overall method results. The papers also emphasize the combination of BN with other machine learning methods like genetic algorithms.

        The examples of NIDS models using Support Vector Machine (SVM) are demonstrated in

      • ,

        The papers point out good generalization of SVM so

        important for detection of new network intrusions and real-time processing because of its lower training and testing time. The SVM capability of accurate classification is mostly used in the hierarchical hybrid models where the SVM plays a role of the final classifier whereas other methods are responsible for preceding dynamic feature extraction and dimension reduction.

        Various types of Arti ficial Neural Networks (ANN) and their utilization in NIDS are addressed in [


        The main points resulting from analysis of the

        papers are: the multilevel preprocessing and feature extraction that come from the usage of emergent deep learning in the form of deep neural networks, and hierarchical ensemble methods that are built upon a set of simple ANN models specialized in detection of individual DoS attack types.

        The application of Self-Organizing Map (SOM) to the NIDS domain is addressed


        Considering intrusion detection tasks, the main SOM advantages are the

        unsupervised learning mode, traffic visualization, parallel processing, and real-time

        224 J. Hrabovsky et al.

        spatially spread the detection complexity among many simple nodes placed in the network infrastructure. Such approach brings today so important scalability.

        3.2 Evaluation Challenges Clear evaluation of so many intrusion detection methods faces several issues related to the dissimilarity

         ], such as different implementation approaches (simulation, emu-

        lation, and real deployment) and different purposes of methods according to their categorization.

        The second reason points out the importance of the appropriate dataset selection because the datasets used for evaluation of current methods are diverse and usually obsolete. Furthermore, the ratio of normal traffic to attacks in these datasets is ques- tionable, while detection methods require balanced datasets to learn all types of traffic under same conditions. Present and publicly available datasets (KDD-99 [



        have these

        shortcomings, which improperly influence test results. On top of that, they do not correspond to the behavior of current real network infrastructures.

        Correct evaluation of achieved detection results requires to consider many issues that raise the requirements on testing techniques.

      • Reliability - the test gives relevant information regardless of the type of tested method.
      • Reproducibility - researchers must be able to repeat the experiments under the same conditions with the same results.
      • Flexibility - the test should provide some variability through the setting of various parameters.
      • Scalability - the test should be applicable in the real environment, i.e., inside of a wide network environment with huge network traffic.
      • Processing - the test should provide the results in clear form. The visualization plays important role in network traffic analysis.

        The trend of NIDS evaluation lies in the utilization of data captured from the real network traffic in order to reflect behavior of modern computer networks.

      4 Discussion

        The conducted analysis and comparison of NIDS methods proposed in various research papers lead to potential improvements and development trends. Some of them (the most interesting in our opinion) are summarized in the following list:

      • Multi-level data preprocessing - Continual data refining allows to extract com- posed domain-specific features that lead to discovery of hidden relations in data. Research areas, such as image and natural language processing, apply gradually deeper model structures to enhance the abstraction level of hidden features, e.g. deep neural and deep belief networks and thus improve the results.
      • Shift from traditional to hybrid methods - The hybrid approach eliminates drawbacks of supervised and unsupervised learning through their cooperation (they complement each other). The specialization allows sub-models to deal with the tasks, where they excel.
      • Automatic feature selection - This task is important step for any machine learning method regardless of its application domain. Feature selection can be seen as individual problem that can be solved through the machine learning, too.
      • Real network traffic for training - Training on real network traffic deals with issues that closely relate to the properties of available network datasets described in
      • Distributed computation - Ensemble method composed of several elementary models enables real-time processing through the parallel computation. At the same time, the network of detectors (cooperative models that provide collective results) simplifies adaptation to the wide complex infrastructure.
      • Graphical format of the method processing and results - The visualization supports better understanding of algorithm principles and its behavior and provides the additional format of outputs.
      • On-line model update - Dynamic environment of current computer networks demands the on-line learning in order to react sufficiently to irregular changes in the network behavior. The on-line adjustment is not focused only to model parameters but also to the dynamic model structure.

        By applying the preceding trends to the method design, we expect the improvement of the overall performance and solutions to problems in current NIDS.

        The paper deals with issues of NIDS according to the growing impact of DoS/DDoS attacks on the quality of network services. Because of the increasing dependency on the permanent availability of computer networks, we pointed out the importance of new method development that should consider here achieved summaries and identified method drawbacks. Both sides of the NIDS domain - network attacks and detection methods - were introduced as the background of the network intrusion detection domain. The paper describes their hierarchical classification with focus on machine learning algorithms applied as the anomaly-based NIDS. Related research papers of four particular methods – BN, SVM, ANN, and SOM – reflecting current research were summarized. As the analyzed papers use different evaluation methods and thus provide different performance outputs, we emphasize reasons of insufficient comparison and challenges related to the unified evaluation of detection methods. The paper identifies common properties of analyzed methods that are responsible for improved perfor- mance. Trends, which influence method design, are finally outlined.

        The paper was written with the goal of providing general overview of NIDS methods with focus on anomaly-based detection approach, the machine learning in particular. Its purpose is the analysis of current research trends in order to identify and

        Trends in Application of Machine Learning 225

      5 Conclusion

        specify further direction. Therefore, as the main contribution of the paper, we are considering the explicit determination of trends that we derived from the analysis of relevant research articles and enumerated in discussion section.


      10. Singh, M.D.: Analysis of host-based and network-based intrusion detection system. Int.


      19. Vijayasarathy, R.: A systems approach to network modelling for DDoS detection using

      Naive Bayes classifier. In: Communication Systems and Networks (COMSNETS). IEEE, January 2011

      20. Kumar, G., Kumar, K.: Design of an evolutionary approach for intrusion detection. Sci.

      World J. 2013, 14 (2013).


      18. Patel, K.K., Buddhadev, B.V.: Machine learning based research for network intrusion

      detection: a state-of-the-art. Int. J. Inf. Netw. Secur. 3(3), 31–50 (2014).


      15. Vijaykumar, B., Vikramkumar, Trilochan: Bayes and Naive Bayes Classifier. arXiv (2014)

      16. Cortes, C., Vapnik, V.: Support-vector networks. Mach. Learn. 20(3), 273–297 (1995).

      . ISSN: 1573-0565 17. Kohonen, T.: The self-organizing map. Proc. IEEE 78(9), 1464–1480 (1990).


      14. Pearl, J.: Fusion, propagation, and structuring in belief networks. Artif. Intell. 29(3), 241–

      288 (1986).


      13. Creech, G., Hu, J.: A semantic approach to host-based intrusion detection systems using

      contiguousand discontiguous system call patterns. IEEE Trans. Comput. (2014).

        12. Gerhards, R.: RFC 5424 - The Syslog Protocol (2009)


      11. Letou, K., Devi, D., Singh, Y.J.: Host-based intrusion detection and prevention system

      (HIDPS). Int. J. Comput. Appl. 69(26), 28–33 (2013).

        J. Comput. Netw. Inf. Secur. 8(8), 41–47 (2014).


      9. Bhattacharyya, D.K., Kalita, J.K.: Network Anomaly Detection: A Machine Learning

      Perspective, p. 366. Chapman and Hall/CRC, Boca Raton (2013). ISBN: 9781466582088


        Auerbach Publications, Boca Raton (2011). ISBN: 9781439839423


      7. Geva, M., Herzberg, A., Gev, Y.: Bandwidth distributed denial of service: attacks and

      defenses. IEEE Secur. Priv. 12(1), 54–61 (2014).

      8. Dua, S., Du, X.: Data Mining and Machine Learning in Cybersecurity, 1st edn, p. 256.

        6. Holmes, D.: 2016 DDoS Attack Trends (2016)


      5. Neustar: The threatscape widens: DDoS aggression and the evolution of IoT risks (2016)

        4. Neustar: Worldwide DDoS Attacks & Protection Report (2016)


      3. Zlomisli, V., Fertalj, K., Vlado, S.: Denial of service attacks : an overview. In: 2014 9th

      Iberian Conference on Information Systems and Technologies (CISTI) (2014).


      2. Handley, M.J., Rescorla, E.: RFC 4732 - Internet Denial-of-Service Considerations, pp. 1–38



      1. Douligeris, C., Mitrokotsa, A.: DDoS attacks and defense mechanisms: classification and

      state-of-the-art. Comput. Netw. 44(5), 643–666 (2004).


      21. Thottan, M.: Anomaly detection in IP networks. IEEE Trans. Signal Process. 51(8), 2191–

      2204 (2003). 226 J. Hrabovsky et al.

        Trends in Application of Machine Learning 227


      22. Alkasassbeh, M., Al-Naymat, G., Hassanat, A.B.A., Almseidin, M.: Detecting distributed

      denial of service attacks using data mining techniques. Int. J. Adv. Comput. Sci. Appl. 7(1), 436–445 (2016).


      23. Chandola, V., Banerjee, A., Kumar, V.: Anomaly detection. ACM Comput. Surv. 41(3), 1–58



      24. Osareh, A., Shadgar, B.: Intrusion detection in computer networks based on machine

      learning algorithms. Ijcsns 8(11), 15 (2008)


      25. Kim, G., Lee, S., Kim, S.: A novel hybrid intrusion detection method integrating anomaly

      detection with misuse detection. Expert Syst. Appl. 41(4), 1690–1700 (2014). . PART 2


      26. Erfani, S.M., Rajasegarar, S., Karunasekera, S., Leckie, C.: High-dimensional and

      large-scale anomaly detection using a linear one-class SVM with deep learning. Pattern Recognit. 58, 121–134 (2016).


      27. She, C., Wen, W., Lin, Z., Zheng, K.: Application-Layer DDOS Detection Based on a

      One-Class Support Vector Machine. Int. J. Netw. Secur. Appl. 9(1), 13–24 (2017).


      28. Alfantookh, A.A.: DoS attacks intelligent detection using neural networks. J. King Saud

      Univ. Comput. Inf. Sci. 18, 31–51 (2006).


      29. Javidi, M.M., Nattaj, M.H.: A new and quick method to detect DoS attacks by neural

      networks. J. Math. Comput. Sci. 6, 85–96 (2013)


      30. Tang, T.A., Mhamdi, L., McLernon, D., Zaidi, S.A.R., Ghogho, M.: Deep learning approach

      for network intrusion detection in software defined networking. In: 2016 International Conference on Wireless Networks and Mobile Communications (WINCOM), pp. 258–263 (2016).


      31. Garcia, M.A., Trinh, T.: Detecting simulated attacks in computer networks using resilient

      propagation artificial neural networks. Polibits 51, 5–10 (2015).


      32. Wei, M., Su, J., Jin, J., Wang, L.: Research on intrusion detection system based on BP neural

      network, vol. 270. LNEE, vol. 1, pp. 657–663 (2014).


      33. Li, J., Liu, Y., Gu, L.: DDoS attack detection based on neural network. In: 2010 2nd

      International Symposium on Aware Computing (ISAC), pp. 196–199 (2010).


      34. Mitrokotsa, A., Douligeris, C.: Detecting denial of service attacks using emergent

      self-organizing maps. In: Proceedings of the Fifth IEEE International Symposium on Signal

        Processing and Information Technology, vol. 2005, pp. 375–380 (2005).



      35. Pan, W., Li, W.: A hybrid neural network approach to the classification of novel attacks for

      intrusion detection. In: Pan, Y., Chen, D., Guo, M., Cao, J., Dongarra, J. (eds.) ISPA 2005. LNCS, vol. 3758, pp. 564–575. Springer, Heidelberg (2005). . ISBN: 978-3-540-32100-2


      36. Wang, C., Yu, H., Wang, H., Liu, K.: SOM-based anomaly intrusion detection system. In:

      Kuo, T.-W., Sha, E., Guo, M., Yang, Laurence T., Shao, Z. (eds.) EUC 2007. LNCS, vol. 4808, pp. 356–366. Springer, Heidelberg (2007). ISBN: 978-3-540-77092-3


      37. Jiang, D., Yang, Y., Xia, M.: Research on intrusion detection based on an improved SOM

      neural network. In: 2009 Fifth International Conference on Information Assurance and Security, pp. 400–403 (2009).


      38. Choksi, K., Shah, B., Ompriya Kale, A.: Intrusion detection system using self organizing

        228 J. Hrabovsky et al.


      39. Kim, M., Jung, S., Park, M.: A distributed self-organizing map for DoS attack detection. In:

      2015 Seventh International Conference on Ubiquitous and Future Networks, pp. 19–22.

        IEEE (2015).


      40. Behal, S., Kumar, K.: Trends in validation of DDoS research. Procedia Comput. Sci. 85, 7–15



      41. Tavallaee, M., Bagheri, E., Lu, W., Ghorbani, A.A.: A detailed analysis of the KDD CUP 99

      data set. In: IEEE Symposium on Computational Intelligence for Security and Defense

        Applications CISDA 2009, no. Cisda, pp. 1–6 (2009).


        42. The CAIDA UCSD ‘DDoS Attack 2007’ Dataset.


        43. DARPA Scalable Network Monitoring (SNM) Program Traffic.


      44. Gogoi, P. Bhuyan, M.H., Bhattacharyya, D.K., Kalita, J.K.: Packet and flow based network

      intrusion dataset. Contemp. Comput., 322–334 (2012). ISBN 978-3-642-32129-0


      Acoustic Signal Classification Algorithm

      for WSN Node in Transport System

        Road transport is a very interesting area where WSN can be used in various ways


        these acoustic signals and, last but not least, to reduce the data needed for the successful classification of the acoustic signal

        . Another goal is to create a new method for classifying

        The set of tasks for the use of acoustic signal analysis in road transport is broad. That’s why our goal is to focus mainly on the classification of different acoustic signals

        As an example, sensing the acoustic emissions in transport, sensing the movement of people in the building or acquisition of meteorological data, or widespread use of wireless sensor networks (WSN) in intelligent buildings.

        the collection of information from many sensors that are suitably divided into large zones. This is why research and development has its natural focus on implementing a wireless sensor into a road network that would serve to monitor and manage the application.

        . Obviously, monitoring and control, eventual traffic control, requires

        Transport system

        R´obert ˇ Zalman

        Neural network ·

        Keywords: Acoustic ·

        In the paper, we focus on the classification of the acoustic

      signal and its characteristic properties, which we use for further process-

      ing of the acoustic signal. Its further processing is ensured that we are

      able to find the carrier frequencies of the selected signal with frequency

      analysis. We use compression methods to reduce the data needed to clas-

      sify acoustic signals. We use neural networks to classify these signals. In

      addition, a method has been proposed to classify acoustic signals that are

      commonly found in transport. The result is the design of a method that is

      able to classify signals characteristic for different environments or differ-

      ent acoustic sources. In the paper, there is a description of the experiment

      that has been carried out for the mentioned purposes. For experiment

      is created evaluation and classification success rate on selected acoustic




      Faculty of Management Science and Informatics,

      University of ˇ Zilina, Univerzitn´ a 8215/1, 010 26 ˇ Zilina, Slovakia


        , Michal Chovanec, Martin Rev´ ak, and J´ an Kapitul´ık


        ( B

      1 Introduction

        230 R. ˇ Zalman et al.

        2 Proposal of Method for Classification

        The proposal for a method for classifying acoustic signals in transport is shown in the Fig.

        With this method, we have succeeded in achieving the greatest success of the acoustic signal classification.

        Fig. 1.

        Block diagram of proposed method for classification.

        The proposed method involves following steps. Using Frequency Analysis, we get the carrier frequencies from an acoustic signal. In experiments, we tested dif- ferent frequency analyzes and procedures, of which they had the best results for frequency filters. Especially because of their rapid response and lower computa- tional demands. By creating a time window, we can preserve the time changes of the signal. The course of acoustic signals is not a one-off event, so it is necessary to preserve their dependence over time (e.g. siren sound). We need to reduce the data we work with, while retaining their carrier information. Data was reduced using the principal component analysis. Data reduction is required for fast pro- cessing and evaluation of data using the neural network. Due to the reduction, it is not necessary for the neural network to be robust. The neural network is of a forward type and uses the Back - Propagation algorithm. Other types of neural networks and algorithms can also be used. For individual blocks of the proposed method, experiments were created to verify individual steps.

        3 Synthesis

        The proposed solution includes filters, normalization, time window, PCA and NN parameters that we explain in detail.

        3.1 Design of Frequency Filter Second order resonant filters, IIR filters (y[n] = x[n] − a  y[n − ] − a  y[n − ]), can be designed with one peak, in its frequency characteristic, according to the coefficients as follows: a


      1 c T vz ),

        = −2rcos(2πf (1)


        Acoustic Signal Classification Algorithm 231

        where f c is the center of the resonant frequency, T vz = /f vz , thus f vz is sampling frequency and r from interval (, ). When r is going to , the bandwidth is narrowed and r can be defined as:

        1 , (2) r = 1 −

        1 + C f where C f is count of filters. Next, we can design a resonant filter so that the resonant peak always has a gain of . by specifying the numerator coefficients as: y[n] = b x[n] + b


        1 x[n − 1] + b 2 x[n − 2] − a 1 y[n − 1] − a 2 y[n − 2]

        where: 2

        (1−r )

        b = ,


        (4) b = 0,

        1 b . 2 = −b

        This way, it is possible to create one filter at the selected frequency. It is also possible to connect many resonant filters in parallel, so each filter resonates at a different frequency and gives its own output at the selected frequency.

        The resonant filters align with the frequency at which they are set. They are able to pass or suppress selected frequencies. These filters are typically described by their resonance frequency and quality factor K, which describes their sharpness.

        We determined the bandwidth for a single filter using the sampling frequency and the required number of filters: F vz

        B = , (5) w C f

        F vz f = j, j = 0, 1, ..., C , (6) cj f

        2C f where B w is bandwidth for one filter, F vz is sampling frequency, C f is count of filters and f cj is the center of the resonant frequency for j-th filter.

      4 Normalization

        It’s necessary to normalize output data from the filters to the interval (−, ) to approve further neural network processing. This normalization is done by fitting using: y = ky F + q, (7) F where y is normalized output, y F is F -th filter output, F


        k = , y −y Fmax Fmin F ∗ k. max (8) q = 1 − y

        232 R. ˇ Zalman et al.

        4.1 Time Window A time window is necessary because it is problematic to determine whether it is a real acoustic signal or just an acoustic anomaly that lasts only a few ms and ended. We can capture this time sequence by using the created time window. This window keeps data in time using certain time steps. We can present these steps as several listed vectors that contain data from different time points in succession.

      5 Principal Component Analysis – PCA

        The primary objective of the PCA is to simplify the description of a group of mutually dependent, i.e., correlated, characters. The method reduces the size of the job, reducing the number of characters with allowed loss of information, which is especially useful for displaying multidimensional data. Individual mea- sured values quite often show a strong correlation. To simplify the analysis and easier evaluation of results, it is appropriate to examine whether it is possible the entire group of variables (i.e., studied the properties of the observed objects) replaced with a single variable or a smaller number of variables that will contain data on nearly the same information as contained original value. This problem can be described as a method of linearly transforming the original characters into new, uncorrelated variables that have more convenient properties and are significantly less.

        The principal component analysis finds the hidden (artificial, latent, non- measurable) quantities, called the main components. The newly created variables are a linear combination of the original variables and they are required to best represent the original variables, in other words, to best explain the variability of the original variables.

        PCA algorithm is: j −

        1 w x

        y j i , (9) = x − i

        =0 where x is input vector and w is a matrix of eigenvectors.

        For past learning we used approximate learning with Oja’s rule. Oja rule that we used to learn PCA algorithm can be written as: j ∆w ji = ηV j (x ji V k w ki ), (10)

        − k =0 where η is speed of learning, V is j-th neuron, N -dimensional input pattern x made of the distribution P (x), w ki are weights of synapsis from neuron k to neuron i.

        In practice, the main component method is used, for example, to effectively recognize human face images. In this case, the main component method reduces the original image space and provides a very sensible extraction of features. The practical task is to identify people according to the chosen biometric feature, such as eye iris or facial features


        Acoustic Signal Classification Algorithm 233

      6 Neural Network Neural networks show interest not only in the professional but also in the public.

        The simulations of these networks surprisingly yield very good results. As has been said, artificial neural networks (hereafter neural networks) are simplified mathematical models of nervous systems of living organisms. They demonstrate the ability of human thinking to learn


        The mathematical model of the neural network is based on the artificial (formal) neuron that we obtain by reformulating the simplified function of the neurophysical neuron into mathematical speech.

        Artificial neuron (here after neuron) has N generally real inputs x  , ..., x n that model dendrites. Inputs are generally rated by real synaptic weights w  , ..., w n that determine their permeability. In accordance with neurophysi- cal motivation, the synaptic weights can be negative, expressing their inhibitory nature.

        The weighted sum of input values represents inner potential of the neuron: n u = w i x i . (11) i



        The value of the internal potential of u after the so-called Threshold value Θ induces the output of the neural y, which model the electric pulse of the axon. The nonlinear increase of the output value y = S(u) at the threshold value of Θ is given, With the S activation function.

        Using a formal modification, the S function will have a zero threshold, and our own threshold of the neuron will be understood as the weight, The bias of another formal input with a constant unit value .

        6.1 Neural Network Learning Teaching the ability of neural networks lies in the ability to change all weights in the network according to appropriate algorithms, unlike biological networks, where the ability to learn is based on the possibility of creating new connections between neurons. Physically, they are therefore both learning ability based on different principles, but not in terms of logic. Algoritmus Back-Propagation Algoritmus Back-Propagation is used in about 80% of all neural network appli- cations. The algorithm itself has three stages: feed-forward spreading of the training pattern input signal, error redistribution and updating of weight values on connections.

        For recognition, we used a feed-forward neural network with tangents hyper- bolic transfer function: n y = tanh( w i x i ), (12) i


        234 R. ˇ Zalman et al.

        for hidden layers and linear transfer function for output layer: n y = w i x i . (13) i



        7 Experiment As suggested by the proposed method on the block schema Fig. by adding the

        principal component analysis to reduce the size of the input matrix to the neural network, we have retained the signal carrier information and even reduced the data needed to successfully identify it. The block diagram of the experiment is shown in the Fig.


        Fig. 2.

        Block diagram of the resulting experiment.

        7.1 Input Database The input database is created by random access to the data from which one time window is filled. The next window is filled in by random access. This is repeated until the input stream is created. Two random input streams are created for testing and training.

        In the experiment, it was necessary to use so-called raw data (data without a header) whose individual parameters are shown in the Table


        Table 1.

        Input data parameters

      Number of channels Single channel (Mono)

      Sampling rate 44100 Hz Coding 32 bit float Endianness Little-endian

        Creating input data consists of randomly selecting the time slots from the individual data shown in the Table


        Acoustic Signal Classification Algorithm 235

      Table 2. Types of input data

      rain.raw, leaf.raw, siren.raw

        3 Types of recorded data, ca. 2 min

      water.raw, white noise.raw, saw.raw 3 Types of recorded data, ca. 2 min

      wind.raw, PassingCars.raw, car.raw 3 Types of recorded data, ca. 2 min

        The algorithm randomly accesses individual files in set, randomly chooses a position and selects a time period (e.g., 1 s). This algorithm passes through individual files if it does not collect a sufficient number of samples (e.g., 60 s). Two files are created in this way. Training file and test file. The input data stream is shown in the Fig.



      Fig. 3. The input stream of data, length 60 s.

        The spectrogram for the given waveform is shown in the Fig.



      Fig. 4. Spectrogram for input stream of 60 s.

        The number of input raw data in the experiment varied to the stage as described above. At the beginning of the experiment, only three types of data were compared: siren, saw, car, later on in the experiment, we were comparing 9 inputs shown in Table

        This was due to the fact that the frequency represen-

        tation of the individual data is diametrically different: the car is represented by low frequencies, the saw also contains higher harmonic components (it manifests almost in the whole frequency spectrum) and the siren contains frequencies in the range of 600–1350 Hz with repetition 12 times per minute.

        Another procedure was to add individual data sets. The last added set of data was the generated white noise, its purpose is to increase the inaccuracy of the whole evaluation system.

        236 R. ˇ Zalman et al.

        7.2 Frequency Filters In our case, we tested different number of resonant filters (Table

        , which pro- cessed and distributed input data in the frequency domain.

        Table 3.

        Success rate with various numbers of resonant filters

      Number of filters Success (Output from neural network)

      180 94% 200 98% 210 55% 220 30%

        The experiment shows that with increasing number of filters, success has grown into a number of filters. With a certain number of filters, success began to decline. The Table

         shows that the best results we achieved with the number of filters , where F vz =  Hz a C f = , bandwidth for each filter ,  Hz.

        7.3 Time Window The filter outputs are regularly stored to create a time window that is further processed. The time window size is variable and can vary from a few ms to several seconds. Example, from  ms to  s. In the experiment, we found that the best results are obtained at the window  s and the data changes every  ms. Input for the time window is a vector with the size of 100 elements, for 4 s time window is output matrix with a size of 2000 elements 20 columns and 100 rows. Time Window Format The result of the filters is stored in the vector every  ms. In this way, the vectors are stored up to the required number of  ( × . =  s).

        7.4 Principal Component Analysis – PCA PCA – The principal component analysis serves us mainly to reduce the two- dimensional space with a large number of elements on a one-dimensional space with a relatively small number of elements. The matrix of  element size (input from the time window  × ) is reduced to a vector of  element size.

        The PCA algorithm was tested primarily on the data of the numbers that were decomposed and re-composed with the error %. This error may seem big. Our goal, however, is not to recompose the original signal, but to build its principal components, which we will continue to process. We can imagine this as a frequency spectrum normalized to interval (−, ) entering the PCA. The output is the principal component vector and their weight for each component.

        Acoustic Signal Classification Algorithm 237

        7.5 Neural Network Using the neural network, we are able to classify input data. Although our app- vz roach reduces input data from the size of  ×  (input F × seconds) to a vector of  elements, the neural network proposed by doesn’t need to be robust to correctly classify the acoustic signal. We manage to run many experiments with different parameters of the system, best result was obtained with follows network topology (Table

        . The proposed neural network contains one hidden

        layer. The network topology is as follows: the input layer –  neurons (PCA output), one hidden layer with  neurons, and the output layer with  neurons, where each belongs to an individual set of data.

        8 Classification Success Rate

        From these experiments, we can claim that the classification is successful for the parameters listed in the Table



      Table 4. Results of classification experiment.

      Number of input elements 176 400

      Number of filters 200 Length of time window 4 s Number of samples in time window

        20 Number of PCA components

        95 Number of hidden layers NN

        1 The number of neurons in the hidden layer 95 Number of classified outputs

        9 Success rate 98% We can say that the % success rate is sufficient for our requirements.

        9 Conclusion

        This paper proposes a method of processing audio signal. The method is suitable for classifying acoustic signal sources when monitoring the emitted sound in road traffic. The results can be further used in traffic flow classification, crossing control, vehicle monitoring, etc. The proposed method focuses on reducing the data needed to recognize the acoustic signal, with no less successful classification.

        An acoustic signal analysis and the possibility of using its properties in clas- sification were also performed. In the experiment, a method for classifying dif- ferent acoustic signals was created. We conducted an experiment that served to implement the proposed method.

        The ability of the proposed method to successfully classify different types of acoustic signal. The positive side of this method is the extension of other acoustic signals that it is capable of recognizing.

        238 R. ˇ Zalman et al.



      1. Astapov, S., Preden, J.S., Suurjaak, E.: A method of real-time mobile vehicle

      identification by means of acoustic noise analysis implemented on an embedded device. In: 2012 13th Biennial Baltic Electronics Conference (BEC), pp. 283,286, 3–5 October 2012

        2. Oleˇsnan´ıkov´

        a, V., P´ uchyov´

        a, J.: Analysis of voice activity detection for implemen- tation into WSN. In: CSIT 2014, Lviv, Ukraine, pp. 75–76 (2014). ISBN 978-617- 607-669-8 3. ˇ Zalman, R., Oleˇsnan´ıkov´ a, V., ˇ Saraf´ın, P., Kapitul´ık, J.: Analysis of acoustic signals in transport systems using WSN. In: TRANSCOM 2015, ˇ Zilina, Slovak Republic, pp. 105–109 (2015). ISBN 978-80-554-1045-6


      4. Hodon, M., ˇ Saraf´ın, P., ˇ Sevˇc´ık, P.: Monitoring and recognition of bird population

      in protected bird territory. In: ISCC 2015, Larnaca, Cyprus, pp. 993–998 (2015).

        ISBN 978-1-4673-7194-0

        5. Oleˇsnan´ıkov´

        a, V., Karpiˇs, O., Chovanec, M., ˇ Saraf´ın, P., ˇ Zalman, R.: Water level monitoring based on the acoustic signal using the neural network. In: FedCSIS 2016, Rzeszow, Poland, pp. 203–206 (2016). ISBN 978-1-4673-8860-3

        6. Chovanec, M., P´ uchyov´

        a, J., H´ udik, M., Kochl´ aˇ n, M.: Universal synchronization algorithm for wireless sensor networks - “FUSA algorithm”. In: FedCSIS 2014, Warsaw, Poland (2014). ISSN 2300–5963

        7. Miˇcek, J., Karpiˇs, O., Oleˇsnan´ıkov´

        a, V., Kochl´ aˇ n, M.: Monitoring of water level based on acoustic emissions. In: ISCC 2015, Larnaca, Cyprus, pp. 988–992 (2015).

        ISBN 978-1-4673-7194-0


      8. Fritsch, L.: Metoda PCA a jej´ı implementace v jazyce C++. ˇ CVUT in



      9. Kr´ atk´ y, M., Skopal, T., Sn´ aˇsel, V.: Efektivn´ı vyhled´ av´ an´ı v kolekc´ıch obr´ azk˚ u

      tv´ aˇr´ı. In: Proceedings of DATAKON 2003, Brno, Czech Republic. Vˇ SB, Ostrava (2003). ISBN 80-210-3215-4.


      10. Jirs´ık, V., Hr´ aˇcek, P.: Neuronov´e s´ıtˇe, expertn´ı syst´emy a rozpozn´ av´ an´ı ˇreˇci, 106

      p. VUT, Brno (2002)

      11. ˇ S´ıma, J., Neruda, R.: Teoretick´e ot´ azky neuronov´ ych s´ıt´ı, 390 p. MATFYZPRESS,

        Author Index Apel, Sebastian

         Lüke, Karl-Heinz Matiasko, Karol

         Von Hollen, Sven Walther, Johannes Wilhelm, Geoffrey

         Tran, Haï

         Thomas, Kevin

         Semm, Arlett Späthe, Steffen

         Revák, Martin Schreyer, Martin Scius-Bertrand, Anna

         Papan, Jozef Piamrat, Kandaraj

         Moravcik, Marek Olešnaníková, Veronika

         Molka-Danielsen, Judith


         Kvet, Michal ,

         Ayaida, Marwane

         Jaccheri, Letizia Karpiš, Ondrej

         Ho, Kim Thoa

         Hertrampf, Florian

        Groß, Rainer Herbin, Michel

        Chudzikiewicz, Jan Erfurth, Christian Fouchal, Hacène ,

         Čechovič, Lukáš

         Bui, Quang Vu


         Bourdy, Emilien

         Bosom, Jérémie

         Žalman, Róbert

Dokumen baru