Handbook of Multimedia for Digital Entertainment and Arts- P1 pot

Handbook of Multimedia for Digital Entertainment
and Arts


Borko Furht
Editor

Handbook of Multimedia
for Digital Entertainment
and Arts

123


Editor
Borko Furht
Department of Computer Science
and Engineering
Florida Atlantic University
777 Glades Road
PO Box 3091
Boca Raton, FL 33431
USA


ISBN 978-0-387-89023-4
e-ISBN 978-0-387-89024-1
DOI 10.1007/978-0-387-89024-1
Springer Dordrecht Heidelberg London New York
Library of Congress Control Number: 2009926305

c Springer Science+Business Media, LLC 2009
All rights reserved. This work may not be translated or copied in whole or in part without the written
permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York,
NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in
connection with any form of information storage and retrieval, electronic adaptation, computer software,
or by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are
not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject
to proprietary rights.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)


Preface

The advances in computer entertainment, multi-player and online games,
technology-enabled art, culture and performance have created a new form of entertainment and art, which attracts and absorbs their participants. The fantastic success
of this new field has influenced the development of the new digital entertainment
industry and related products and services, which has impacted every aspect of our
lives.
This Handbook is carefully edited book – authors are 88 worldwide experts in
the field of the new digital and interactive media and their applications in entertainment and arts. The scope of the book includes leading edge media technologies and
latest research applied to digital entertainment and arts with the focus on interactive
and online games, edutainment, e-performance, personal broadcasting, innovative
technologies for digital arts, digital visual and auditory media, augmented reality,
moving media, and other advanced topics. This Handbook is focused on research
issues and gives a wide overview of literature.
The Handbook comprises of five parts, which consist of 33 chapters. The first
part on Digital Entertainment Technologies includes articles dealing with personalized movie, television related media, and multimedia content recommendations,
digital video quality assessments, various technologies for multi-player games, and

collaborative movie annotation. The second part on Digital Auditory Media focuses
on articles on digita music management and retrieval, music distribution, music
search and recommendation, and automated music video generation. The third part
on Digital Visual Media consists of articles on live broadcasts, digital theater, video
browsing, projector camera systems, creating believable characters, and other aspects of visual media.
The forth part on Digital Art comprises articles that discuss topics such as information technology and art, augmented reality and art, creation process in digital art,
graphical user interface in art, and new tools for creating arts. The part V on Culture
of New Media consists of several articles dealing with interactive narratives, discussion on combining digital interactive media, natural interaction in intelligent spaces,
and social and interactive applications based on using sound-track identification.
With the dramatic growth of interactive digital entertainment and art applications, this Handbook can be the definitive resource for persons working in this field
as researchers, scientists, programmers, and engineers. The book is intended for a

v


vi

Preface

wide variety of people including academicians, animators, artists, designers, developers, educators, engineers, game designers, media industry professionals, video
producers, directors and writers, photographers and videographers, and researchers
and graduate students. This book can also be beneficial for business managers, entrepreneurs, and investors. The book can have a great potential to be adopted as a
textbook in current and new courses on Media Entertainment.
The main features of this Handbook can be summarized as:
The Handbook describes and evaluates the current state-of-the-art in multimedia
technologies applied in digital entertainment and art.
It also presents future trends and developments in this explosive field.
Contributors to the Handbook are the leading researchers from academia and
practitioners from industry.
I would like to thank the authors for their contributions. Without their expertise

and effort this Handbook would never come to fruition. Springer editors and staff
also deserve our sincere recognition for their support throughout the project.
Borko Furht
Editor-in-Chief
Boca Raton, 2009


Contents

Part I DIGITAL ENTERTAINMENT TECHNOLOGIES
1

Personalized Movie Recommendation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
George Lekakos, Matina Charami, and Petros Caravelas

3

2

Cross-category Recommendation for Multimedia Content . . . . . . . . . . . . .
Naoki Kamimaeda, Tomohiro Tsunoda, and Masaaki Hoshino

27

3

Semantic-Based Framework for Integration
and Personalization of Television Related Media . . . . . . . . . . . . . . . . . . . . . . . .
Pieter Bellekens, Lora Aroyo, and Geert-Jan Houben

59

4

Personalization on a Peer-to-Peer Television System . . . . . . . . . . . . . . . . . . . .
Jun Wang, Johan Pouwelse, Jenneke Fokker, Arjen P. de Vries,
and Marcel J.T. Reinders

91

5

A Target Advertisement System Based on TV Viewer’s
Profile Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Jeongyeon Lim, Munjo Kim, Bumshik Lee, Munchurl Kim,
Heekyung Lee, and Han-kyu Lee

6

Digital Video Quality Assessment Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Anush K. Moorthy, Kalpana Seshadrinathan,
and Alan C. Bovik

7

Countermeasures for Time-Cheat Detection in Multiplayer
Online Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
Stefano Ferretti

8

Zoning Issues and Area of Interest Management
in Massively Multiplayer Online Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Dewan Tanvir Ahmed and Shervin Shirmohammadi
vii


viii

9

Contents

Cross-Modal Approach for Karaoke Artifacts Correction . . . . . . . . . . . . . 197
Wei-Qi Yan and Mohan S. Kankanhalli

10 Dealing Bandwidth to Mobile Clients Using Games . . . . . . . . . . . . . . . . . . . . . 219
Anastasis A. Sofokleous and Marios C. Angelides
11 Hack-proof Synchronization Protocol for Multi-player
Online Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
Yeung Siu Fung and John C.S. Lui
12 Collaborative Movie Annotation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Damon Daylamani Zad and Harry Agius
Part II DIGITAL AUDITORY MEDIA
13 Content Based Digital Music Management and Retrieval . . . . . . . . . . . . . . 291
Jie Zhou and Linxing Xiao
14 Incentive Mechanisms for Mobile Music Distribution . . . . . . . . . . . . . . . . . . 307
Marco Furini and Manuela Montangero
15 Pattern Discovery and Change Detection of Online Music
Query Streams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327

Hua-Fu Li
16 Music Search and Recommendation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
Karlheinz Brandenburg, Christian Dittmar, Matthias Gruhne,
Jakob Abeòer, Hanna Lukashevich, Peter Dunker, Daniel
Gă rtner, Kay Wolter, Stefanie Nowak, and Holger Grossmann
a
17 Automated Music Video Generation Using Multi-level
Feature-based Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
Jong-Chul Yoon, In-Kwon Lee, and Siwoo Byun
Part III DIGITAL VISUAL MEDIA
18 Real-Time Content Filtering for Live Broadcasts
in TV Terminals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
Yong Man Ro and Sung Ho Jin
19 Digital Theater: Dynamic Theatre Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
Sara Owsley Sood and Athanasios V. Vasilakos
20 Video Browsing on Handheld Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
Wolfgang Hă rst
u


Contents

ix

21 Projector-Camera Systems in Entertainment and Art . . . . . . . . . . . . . . . . . . 471
Oliver Bimber and Xubo Yang
22 Believable Characters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 497
Magy Seif El-Nasr, Leslie Bishko, Veronica Zammitto,
Michael Nixon, Athanasios V. Vasiliakos, and Huaxin Wei
23 Computer Graphics Using Raytracing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529

Graham Sellers and Rastislav Lukac
24 The 3D Human Motion Control Through Refined Video
Gesture Annotation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 551
Yohan Jin, Myunghoon Suk, and B. Prabhakaran
Part IV DIGITAL ART
25 Information Technology and Art: Concepts and State
of the Practice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
Salah Uddin Ahmed, Cristoforo Camerano, Luigi Fortuna,
Mattia Frasca, and Letizia Jaccheri
26 Augmented Reality and Mobile Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593
Ian Gwilt
27 The Creation Process in Digital Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601
Ad´ rito Fernandes Marcos, Pedro S´ rgio Branco,
e
e
and Nelson Troca Zagalo
28 Graphical User Interface in Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 617
Ian Gwilt
29 Storytelling on the Web 2.0 as a New Means
of Creating Arts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623
Ralf Klamma, Yiwei Cao, and Matthias Jarke
Part V CULTURE OF NEW MEDIA
30 A Study of Interactive Narrative from User’s Perspective . . . . . . . . . . . . . . 653
David Milam, Magy Seif El-Nasr, and Ron Wakkary
31 SoundScapes/Artabilitation – Evolution of a Hybrid
Human Performance Concept, Method & Apparatus
Where Digital Interactive Media, The Arts, &
Entertainment are Combined . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 683
A.L. Brooks

x

Contents

32 Natural Interaction in Intelligent Spaces: Designing
for Architecture and Entertainment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713
Flavia Sparacino
33 Mass Personalization: Social and Interactive Applications
Using Sound-Track Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 745
Michael Fink, Michele Covell, and Shumeet Baluja
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 765


Contributors

Jakob Adesser
Fraunhofer Institute, Ilmenau, Germany
Harry Agius
Brunel University, Uxbridge, United Kingdom
Dewan Tanvir Ahmed
University of Ottawa, Ottawa, Canada
Salah Uddin Ahmed
Norwegian University of Science and Technology, Norway
Marios Angelides
Brunel University, Uxbridge, United Kingdom
Lora Aroyo
Eindhoven University of Technology, Eindhoven, The Netherlands
Shumeet Baluja
Google, Mountain View, CA, USA

Pieter Bellekens
Eindhoven University of Technology, Eindhoven, The Netherlands
Oliver Bimber
Bauhaus University Weimar, Germany
Leslie Bishko
Simon Fraser University, Vancouver, Canada
Alan Bovik
University of Texas at Austin, Austin, Texas, USA
Karlheinz Bradenburg
Fraunhofer Institute, Ilmenau, Germany
Pedro Sergio Branco
Computer Graphics Center, Guimaraes, Portugal

xi


xii

Antony Brooks
Aalborg University, Esbjerg, Denmark
Siwoo Byun
Anyang University, Anyang, Korea
Yiwei Cao
Technical University of Aachen, Aachen, Germany
Cristoforo Camerano
University of Catania, Italy
Petros Caravelas
Athens University of Economics and Business, Athens, Greece
Matina Charami
Athens University of Economics and Business, Athens, Greece

Michele Covell
Google, Mountain View, CA, USA
Christian Ditmar
Fraunhofer Institute, Ilmenau, Germany
Peter Dunker
Fraunhofer Institute, Ilmenau, Germany
Magy Seif El-Naser
Simon Fraser University, Vancouver, Canada
Stefano Ferretti
University of Bologna, Bologna, Italy
Michael Fink
The Hebrew University of Jerusalem, Israel
Jennele Fokker
Delft University of Technology, Delft, The Netherlands
Luigi Fortuna
University of Catania, Italy
Mattia Frasca
University of Catania, Italy
Yeung Siu Fung
The Chinese University of Hong Kong, Ma Liu Shui, China
Marco Furini
University of Modena and Reggio Emilia, Italy
Daniel Gartner
Fraunhofer Institute, Ilmenau, Germany

Contributors


Contributors

Holger Grossmann
Fraunhofer Institute, Ilmenau, Germany
Matthias Gruhne
Fraunhofer Institute, Ilmenau, Germany
Ian Gwilt
University of Technology, Sydney, Australia
Masaki Hoshino
Sony Corporation, Tokyo, Japan
Geert-Jan Houben
Eindhoven University of Technology, Eindhoven, The Netherlands
Wolfgang Huerst
Utrecht University, Utrecht, The Netherlands
Letizia Jaccheri
Norwegian University of Science and Technology, Norway
Matthias Jarke
Technical University of Aachen, Aachen, Germany
Subng Ho Jin
Information and Communications University, Deajon, Korea
Yohan Jin
University of Texas at Dallas, Texas, USA
Naoki Kamimaeda
Sony Corporation, Tokyo, Japan
Mohan Kankanhalli
National University of Singapore, Singapore
Munchurl Kim
Information and Communication University, Daejeon, Korea
Munjo Kim
Information and Communication University, Daejeon, Korea
Ralf Klamma
Technical University of Aachen, Aachen, Germany

Bumshik Lee
Information and Communication University, Daejeon, Korea
Han-kyu Lee
Electronics and Telecommunications Research Institute, Deajeon, Korea
Heekyung Lee
Electronics and Telecommunications Research Institute, Deajeon, Korea

xiii


xiv

In-Kwoon Lee
Yonsei University, Seoul, Korea
George Lekakos
Athens University of Economics and Business, Athens, Greece
Hua-Fu Li
Kainan University, Taoyuan, Taiwan
Jeongyeon Lim
Information and Communication University, Daejeon, Korea
John C.S. Lui
The Chinese University of Hong Kong, Ma Liu Shui, China
Rastislav Lukac
Epson Canada Ltd., Toronto, Canada
Hanna Lukashevich
Fraunhofer Institute, Ilmenau, Germany
Aderito Fernnades Marcos
University of Minho, Guimaraes, Portugal
David Milam
Simon Fraser University, Surrey, Canada

Manuela Montangero
University of Modena and Reggio Emilia, Italy
Anush K. Moorthy
University of Texas at Austin, Austin, Texas, USA
Michael Nixon
Simon Fraser University, Vancouver, Canada
Stefanie Nowak
Fraunhofer Institute, Ilmenau, Germany
Johan Pouwelse
Delft University of Technology, Delft, The Netherlands
B. Prabhakaran
University of Texas at Dallas, Texas, USA
Marcel Reinders
Delft University of Technology, Delft, The Netherlands
Yong Man Ro
Information and Communication University, Deajon, Korea
Graham Sellers
Advanced Micro Devices, Orlando, Florida, USA

Contributors


Contributors

Kalpana Seshadrinathan
University of Texas at Austin, Austin, Texas, USA
Shervin Shirmohammadi
University of Ottawa, Ottawa, Canada
Anastas Sofokleus
Brunel University, Uxbridge, United Kingdom

Sara Owsley Sood
Pomona College, Claremont, CA, USA
Flavia Sparacino
Sensing Places and MIT, Santa Monica, CA, USA
Myunghoon Suk
University of Texas at Dallas, Texas, USA
Tomohiro Tsunoda
Sony Corporation, Tokyo, Japan
Arthanasios Vasiliakos
University of Peloponnese, Nauplion, Greece
Arjen de Vries
CWI, Amsterdam, The Netherlands
Ron Wakkary
Simon Fraser University, Surrey, Canada
Jun Wang
Delft University of Technology, Delft, The Netherlands
Huaxin Wei
Simon Fraser University, Vancouver, Canada
Kay Wolter
Fraunhofer Institute, Ilmenau, Germany
Linxing Xiao
Tsinghua University, Beijing, China
Wei-Qi Yan
Queen’s University of Belfast, Belfast, UK
Xubo Yang
Shanghai Jiao Tong University, Shanghai, China
Jong-Chul Yoon
Yonsei University, Seoul, Korea
Damon Daylamani Zad
Brunel University, Uxbridge, United Kingdom

xv


xvi

Nelson Troca Zagalo
University of Minho, Braga, Portugal
Veronica Zammitto
Simon Fraser University, Vancouver, Canada
Jie Zhou
Tsinghua University, Beijing, China

Contributors


Part I

DIGITAL ENTERTAINMENT
TECHNOLOGIES


Chapter 1

Personalized Movie Recommendation
George Lekakos, Matina Charami, and Petros Caravelas

Introduction
The vast amount of information available on the Internet, coupled with the diversity
of user information needs, have urged the development of personalized systems that

are capable of distinguishing one user from the other in order to provide content, services and information tailored to individual users. Recommender Systems (RS) form
a special category of such personalized systems and aim to predict user’s preferences
based on her previous behavior. Recommender systems emerged in the mid-90’s and
since they have been used and tested with great success in e-commerce, thus offering
a powerful tool to businesses activating in this field by adding extra value to their
customers. They have experienced a great success and still continue to efficiently
apply on numerous domains such as books, movies, TV program guides, music,
news articles and so forth.
Tapestry [1], deployed by Xerox PARC, comprises a pioneer implementation in
the field of recommender systems and at the same time, it was the first to embed
human judgment in the procedure of producing recommendations. Tapestry was an
email system capable to manage and distribute electronic documents utilizing the
opinion of users that have already read them. Other popular recommender systems
that followed are Ringo [2] for music pieces and artists, Last.fm as a personalized
internet radio station, Allmusic.com as a metadata database about music genres,
similar artists and albums, biographies, reviews, etc, MovieLens [3] and Bellcore
[4] for movies, TV3P [5], pEPG [6] and smart EPG [7] as program guides for digital
television (DTV), GroupLens [8, 9] for news articles in Usenet and Eigentaste on
Jester database as a joke recommender system. Nowadays, Amazon.com [10] is the
most popular and successful example of applying recommender systems in order to
provide personalized promotions for a plethora of goods such as books, CDs, DVDs,
toys, etc.
G. Lekakos ( ), M. Charami, and P. Caravelas
ELTRUN, the e-Business Center, Department of Management Science and Technology,
Athens University of Economics and Business, Athens, Greece
e-mail: ; ;
B. Furht (ed.), Handbook of Multimedia for Digital Entertainment and Arts,
DOI 10.1007/978-0-387-89024-1 1, c Springer Science+Business Media, LLC 2009

3

4

G. Lekakos et al.

Now more than ever, the users continuously face the need to find and choose
items of interest among many choices. In order to realize such a task, they usually
need help to search and explore or even reduce the available options. Today, there are
thousands of websites on the Internet collectively offering an enormous amount of
information. Hence, even the easiest task of searching a movie, a song or a restaurant
may be transformed to a difficult mission. Towards this direction, search engines
and other information retrieval systems return all these items that satisfy a query,
usually ranked by a degree of relevance. Thus, the semantics of search engines is
characterized by the “matching” between the posted query and the respective results. On the contrary, recommender systems are characterized by features such as
“personalized” and “interesting” and hence greatly differentiate themselves form
information retrieval systems and search engines. Therefore, recommender systems
are intelligent systems that aim to personally guide the potential users inside the
underlying field.
The most popular recommendation methods are collaborative filtering (CF) and
content-based filtering (CBF). Collaborative filtering is based on the assumption
that users who with similar taste can serve as recommenders for each other on unobserved items. On the other hand, content-based filtering considers the previous
preferences of the user and upon them it predicts her future behavior. Each method
has advantages and shortcomings of its own and is best applied in specific situations.
Significant research effort has been devoted to hybrid approaches that use elements
of both methods to improve performance and overcome weak points.
The recent advances in digital television and set-top technology with increased
storage and processing capabilities enable the application of recommendation technologies in the television domain. For example products currently promoted through
broadcasted advertisements to unknown recipients may be recommended to specific
viewers who are most likely to respond positively to these messages. In this way

recommendation technologies provide unprecedented opportunities to marketers
and suppliers with the benefit of promoting goods and services more effectively
while reducing viewers’ advertising clutter caused by the large amount of irrelevant
messages [11]. Moreover, the large number of available digital television channels
increases the effort required to locate content, such as movies and other programs,
that it is most likely to match viewe’s interests. The digital TV vendors do recognize
this as a serious problem, and they are now offering personalized electronic program
guides (EPGs) to help users navigate this digital maze [12].
This article proposes a movie recommender system, named MoRe, which follows a hybrid approach that combines content-based and collaborative filtering.
MoR’s performance is empirically evaluated upon the predictive accuracy of the
algorithms as well as other important indicators such as the percentage of items that
the system can actually predict (called prediction coverage) and the time required
for generating predictions. The remainder of this article is organized as follows. The
next section is devoted to the fundamental background of recommender systems
describing the main recommendation techniques along with their advantages and
limitations. Right after, we illustrate the MoRe system overview and in the section


1 Personalized Movie Recommendation

5

following, we describe in detail the algorithms implemented. The empirical evaluation results are then presented, while the final section provides a discussion about
conclusions and future research.

Background Theory
Recommender Systems
As previously mentioned, the objective of recommender systems is to identify which
of the information items available are really interesting or likable to individual users.
The original idea underlying these systems is based on the observation that people

very often rely upon opinions and recommendations from friends, family or associates to make selections or purchase decisions. Motivated by this “social” approach,
recommender systems produce individual recommendations as an output or have the
effect of guiding the user in a personalized way to interesting or useful objects in a
large space of possible options [13].
Hence, recommender systems aim at predicting a user’s future behavior based on
her previous choices and by relying on features that implicitly or explicitly imply
preferences. As shown in Figure 1, the recommendation process usually takes user
ratings on observed items and/or item features as input and produces the same output
for unobserved items.
Many approaches have been designed, implemented and tested on how to process
the original input data and produce the final outcome. Still, two of them are the most
dominant, successful and widely accepted: collaborative filtering and content-based
filtering. Collaborative filtering is the technique that maximally utilizes the “social”
aspect of recommender systems, as similar users, called neighbors, are used in order
to generate recommendations for the target user. On the other hand, content-based
filtering analyses the content of the items according to some features depending on
the domain in order to profile the users according to their preferences.
These two fundamental approaches are presented in a great detail in the following
subsections. Next, we describe some other alternative techniques used in producing personalized recommendations. We continue by realizing comparative observations among all aforementioned techniques, underlying the strengths and the
shortcomings of each, thus driving the need of combining them in forming hybrid recommender systems. Hybrids form the last subsection of the recommender
systems background theory.

Fig. 1 A high level representation of a recommender system


6

G. Lekakos et al.

Collaborative Filtering

Collaborative filtering comprises the most popular and widely used approach for
generating recommendations [14]. It filters and evaluates items utilizing other people’ tastes and attitudes. It operates upon the assumption that users who have
exhibited similar behavior in the past can serve as recommenders for each other
on unobserved items. Thus, while the term collaborative filtering has become popular since the last decade, its algorithmic behavior originates from something that
people use to do centuries no; exchange views and opinions.
According to collaborative filtering, a user’s behavior consists of her preferences
to products or services. The idea is to trace relationships or similarities between
the target user and the remaining users in the database, aggregate the similar users’
preferences and use them as a prediction for the target user. As a result, users that
seem to prefer and choose common items are identified to have similar purchasing
behavior and belong to a neighborhood. The user of a specific neighborhood may
receive recommendations from her neighborhood for items that she has not bought,
used or experienced in the past with a great possibility of satisfaction as neighbors
are characterized from common taste.
Collaborative filtering consists of four fundamental steps.
1.
2.
3.
4.

Data collection – Input space
Neighbors similarity measurement
Neighbors selection
Recommendations generation

The first step is an independent one that relates to different alternative ways of collecting the input data for the algorithm, while the rest three describe the algorithmic
approach itself.
Data Collection – Input Space
The input space for collaborative filtering may be summarized in a table, called
user ritem matrix, where users form the rows and items form the columns, while

each cell Cij of the matrix is filled by the degree of satisfaction of the ith user for
the j th item. The degree of satisfaction is usually depicted in the form of ratings
from users to items. Ratings can be continuous or discrete in a specific scale, i.e.
ranging from 0 to 100 with real numbers or from 1 to 5 with natural numbers respectively. They can alternatively be provided in a dual representation, i.e. “thumbs
p
up”, 1 or imply that the item was liked and “thumbs down”, 0 or X entail that the
item was not liked. Rating scale chosen depends on the domain of the application.
Nevertheless, the most common rating scale being used is the one that ranges from
1 to 5, with 1 denoting totally unpleasantness and 5 denoting absolute satisfaction.
The empty cells in the user item matrix imply that the user has not yet evaluated
the specific item.
Besides the representation, rating also varies according to the way that it is collected. There exist two different ways of collecting users’ ratings: explicitly and


1 Personalized Movie Recommendation

7

implicitly. Explicit rating refers to a user consciously expressing her preference
usually in a discrete numerical scale. The user evaluates an item and assigns it a
rate according to the scale used. On the other hand, implicit rating refers to interpreting user behavior or selections to impute a vote or preference. It can be based
on browsing data (for example in Web applications), purchase history (for example
in online or traditional stores) or other types of information access patterns.
Explicit rating is much more accurate reflecting more precisely each user’s taste
(as long as users provide consistent ratings), but at the same time, it is much more
difficult to collect from all users and for a large percentage of the offered items.
Moreover, most users usually rate items that they liked and avoid to deal with the
uninteresting ones. Thus, the user item matrix is generally filled in with positive
votes lacking a sufficient amount of negative ones. On the contrary, implicit rating
may not always reflect the reality, since users were not asked directly, and in some

cases, it can also be misleading (e.g. the interpretation of time spent in a website
in the case that the user is idle or has left her computer). Nevertheless, the biggest
advantage of implicit rating is the fact that it relieves the user from examining and
evaluating an item, it is usually based on positive preferences (thus avoiding the
lack of negative ratings) and it manages to continuously collect input data as users
interact with the system. No matter what the nature of the input space is, acquiring
users’ ratings for previously experienced objects comprises the fundamental initial
step for collaborative filtering.
Neighbors Similarity Measurement
Collaborative filtering approaches can be distinguished into two major classes:
model-based and memory-based [15]. Model-based methods develop and learn a
model, which is applied upon the target user’s ratings to make predictions for unobserved items. Two widely used probabilistic models are Bayesian classifier and
Bayesian network with decision trees. On the other hand, memory-based methods
operate upon the entire database of users to find the closest neighbors of the target
and weight their recommendation according to their similarities. The fundamental
algorithm of the memory-based class is the nearest neighbor (denoted as NN), which
is considered as one of the most effective collaborative filtering approaches.
Weighting the neighbors’ recommendations implies defining and calculating the
distance between the target user and her neighbors. This distance may represent
either the correlation or the similarity among all users. A typical measure of correlation is the Pearson correlation coefficient, which indicates the degree of linear
correlation between two variables. In collaborative filtering, it is applied on the
items rated in common by two users. Other popular correlation measures are the
Spearman rank correlation, which is similar to Pearson but calculates the correlation between ranked lists instead of ratings, and the mean-squared difference, which
emphasizes the bigger distances among ratings instead of the small ones. For further details about correlation measures and examples of some systems in where they
were applied, refer to [16].


8

G. Lekakos et al.

On the other hand, the similarity is usually calculated using vectors, the so-called
similarity vectors. In the field of information retrieval, the similarity between two
documents is often measured by treating each document as a vector of word frequencies and computing the cosine of the angle formed by the two frequency vectors.
Adopting this formalism to collaborative filtering, users take the role of the documents, items take the role of words and votes take the role of word frequencies. Note
that in this case, observed votes indicate a positive preference, there is no role for
negative votes and unobserved items receive a zero vote. [15] provides an extensive
description of similarity vectors.
Neighbors Selection
Having assigned weights to users, the next step is to decide which users will be selected and used in the recommendations generation process for the target user. In
other words, select the users that will form the target user’s neighborhood. Theoretically, we could consider all users as neighbors with the closest ones contributing
more and with the more distant ones contributing less in the generation of recommendations. However, real commercial recommender systems deal with thousands
to millions of users, and hence the approach of considering all users in the neighborhood is infeasible in terms of real time response. Thus, the system should select
a subset of users that best form the neighborhood in order to decrease the computational cost and guarantee acceptable response times.
Two techniques have been employed in recommender systems: the thresholdbased selection and the approach of k nearest neighbors (denoted as k–NNs). The
former technique selects as neighbors those users whose correlation or similarity to
the target user exceeds a certain threshold value. Therefore, we may select neighbors
to be over 70% similar to the target user (a high percentage in real online systems).
On the contrary, the latter technique selects a predefined number k of best neighbors
to form target user’s neighborhood. Thus, for example, we may select the 10-best
neighbors of the target user.
The threshold-based approach may form quite reliable neighborhoods, since it
can guarantee that the neighbors of the target user will be, for example, over 70%
similar to her. However, in real applications, the diversity of the users is very high.
In such cases, there exist great possibilities of not even forming neighbors for
some users and as a result, not generating recommendations for those users. On
the other hand, the k nearest neighbors technique always guarantees the formation
of a neighborhood with the cost of possible decrease of results’ quality, since the
distance between the target user and her neighbors may be actually very big.
There is no panacea in choosing one of the aforementioned techniques; rather

the selection of the proper one depends every time on the underlying application
domain. In both cases, evaluating the appropriate values for the threshold or for the
size of the neighborhood is of vital importance. These values depend on the nature
and the size of the input data and are experimentally calculated.


1 Personalized Movie Recommendation

9

Recommendations Generation
As soon as we have assigned weights to users and selected the ones that will serve
as neighbors, we are ready to create predictions for the target user. The recommendation for a new item for the target user is based on the weighted average of her
neighbors’ ratings, weighted by their similarity to the target user. The recommendation generated is normalized in order for its rating to fall in the very same range
used for all items in the domain.

Content-based Filtering
Content-based filtering makes predictions upon the assumption that a user’s previous
preferences or interests are reliable indicators for her future behavior. This approach
requires that the items are described by features, and is typically applied upon textbased documents with predefined format or in domains with structured data, where
the extraction of features that uniformly characterize the data is easy [17].
Text documents are semi-structured data, since they do not consist of specific
predefined words. The application of content-based filtering in semi-structured data
adopts much work from the text information retrieval and the natural language
processing fields, such as representing the documents as vectors and measure the
similarity between those vectors, utilizing attributes and characteristics of the natural language [18]. Applying content-based filtering in unstructured raw data, like
multimedia, proves to be a very interesting and useful task, though a very challenging one that requires a lot of research [19]. Therefore, the majority of developed
content-based recommender systems target on textually described items, like books,
articles and TV programs. Even in the case of music and movies, recommendation
techniques mostly apply on extracted textual features such as title, genre, etc and

not on the multimedia itself. Besides, most on-line music databases today, such as
Napster and mp3.com and movie databases, such as IMDb, rely on file names or text
labels to do searching and indexing, using traditional text processing techniques.
Summarizing, content-based filtering can be applied in book recommender systems using features such as title, author, theme, or even a short summary or
description (if available). In the same sense, it can be applied in a large range of
other application domains, such as recommending movies (with features such as
title, actors, director, genre, plot description), TV programs (with features such as
name, type, presenter, hour), music (with features such as title, artist, album, genre),
restaurants (with features such as name, cuisine, service, cost, place), and so forth.
TF-IDF (term-frequency times inverse document-frequency) and Information
Gain are two metrics commonly applied in content-based filtering [20, 21]. They
are statistical measures used to evaluate how important a word is to a document in
a collection or corpus. The importance increases proportionally to the number of
times a word appears in the document but is offset by the frequency of the word in
the collection. The intuition behind these metrics is that the terms with the highest


10

G. Lekakos et al.

weight occur more often in the current document than in the other documents of
the collection, and therefore are more central to the topic of this document. On the
other hand, terms that are very frequent to all documents in the collection provide
no particular descriptive power for the current document.
Creating a model of the user’s preference from the user history is a form of
classification learning. Such algorithms are the key component of content-based
recommendation systems, because they learn a function that models each user’s interests. Given a new item and the user history, the function makes a prediction on
whether the user would be interested in the item. Some of the most popular algorithms are traditional machine learning algorithms designed to work on structured
data, while other algorithms are designed to work in high dimensional spaces and

do not require a pre-processing step of feature selection. [22] reviews a number of
such well-known classification learning algorithms.

Other Approaches
Apart from the two previously mentioned approaches of collaborative filtering and
content-based filtering, there exist some more alternative techniques that have been
deployed in recommender systems. Some of the most popular ones include the collection of demographic data, the use of a utility function, the creation of a knowledge
model, and the utilization of well-known data mining techniques.
Demographic recommender systems aim to categorize the user based on personal
attributes and make recommendations based on demographic classes. In most of
these systems, demographic data for user categorization are collected through the
interaction of the user with the system using questionnaires, short surveys, dialogue
prompts, etc. In other systems, machine learning is used to arrive at a classifier
based on demographic data. Demographic techniques form ‘people-to-people’ correlations like collaborative ones, but use different data. The benefit of a demographic
approach is that it may not require a history of the type needed by collaborative and
content-based techniques.
Utility-based and knowledge-based recommenders match user’s need with a set
of options available. Utility-based recommenders make suggestions based on a computation of the utility of each object for the user. Of course, the central problem is
how to create a utility function for each user. The benefit of utility-based recommendation is that it can factor non-product attributes, such as vendor reliability and
product availability into the utility computation.
Knowledge-based recommendation attempts to suggest objects based on inferences about a user’s needs and preferences. Knowledge-based approaches have
knowledge about how a particular item meets a particular user need and can therefore reason about the relationship between a need and a possible recommendation.
Data mining has been applied with a great success in the field of grocery retailing.
It usually refers to the automated extraction of implicit but useful information from
large databases. Several data mining techniques, such as classification, clustering