Recommender Systems Handbook (eBook)

Francesco Ricci is associate professor at the faculty of computer science, Free University of Bozen-Bolzano, Italy. His current research interests include recommender systems, intelligent interfaces, mobile systems, machine learning, case-based reasoning, and the applications of ICT to Tourism. He is in the editorial board of Journal of Information Technology and Tourism and he is member of ACM and IEEE. F. Ricci is also member of the steering committee of the ACM Conference on Recommender Systems.Lior Rokach is assistant professor at the Department of Information System Engineering at Ben-Gurion University. He is a recognized expert in intelligent information systems and has held several leading positions in this field. His main areas of interest are Data Mining, Pattern Recognition, and Recommender Systems. Dr. Rokach is the author of over 70 refereed papers in leading journals, conference proceedings and book chapters. In addition he has authored six books and edited three others books.Bracha Shapira is assistant professor at the Department of Information Systems Engineering at Ben-Gurion University, Beer-Sheva, Israel. Her current research interests include recommender systems, information retrieval, personalization, user modelling, and social networks. She leads research projects at the Deutsche telekom Laboratories at Ben-Gurion University and is a member of ACM and IEEE.Paul Kantor is Professor of Information Science in the School of Communication and Information at Rutgers University, with additional appointments in the Faculty of Computer Science and the RUTCOR Center for Operations Research. His interests are in collaborative information finding, text classification, and text or imaging indexing and retrieval. He is a Fellow of the American Association for the Advancement of Science, and a member of the ACM, IEEE and ASIST, and his research is supported by the US NSF and Department of Homeland Security, and other agencies.

Preface 6
Contents 8
List of Contributors 1
Chapter 1 Introduction to Recommender Systems Handbook 28
1.1 Introduction 28
1.2 Recommender Systems Function 31
1.3 Data and Knowledge Sources 34
1.4 Recommendation Techniques 37
1.5 Application and Evaluation 41
1.6 Recommender Systems and Human Computer Interaction 44
1.6.1 Trust, Explanations and Persuasiveness 45
1.6.2 Conversational Systems 46
1.6.3 Visualization 48
1.7 Recommender Systems as a Multi-Disciplinary Field 48
1.8 Emerging Topics and Challenges 50
1.8.1 Emerging Topics Discussed in the Handbook 50
1.8.2 Challenges 53
References 56
Part IBasic Techniques 63
Chapter 2 Data Mining Methods for RecommenderSystems 64
2.1 Introduction 64
2.2 Data Preprocessing 65
2.2.1 Similarity Measures 66
2.2.2 Sampling 67
2.2.3 Reducing Dimensionality 69
2.2.3.1 Principal Component Analysis 69
2.2.3.2 Singular Value Decomposition 70
2.2.4 Denoising 72
2.3 Classification 73
2.3.1 Nearest Neighbors 73
2.3.2 Decision Trees 75
2.3.3 Ruled-based Classifiers 76
2.3.4 Bayesian Classifiers 77
2.3.5 Artificial Neural Networks 79
2.3.6 Support Vector Machines 81
2.3.7 Ensembles of Classifiers 83
2.3.8 Evaluating Classifiers 84
2.4 Cluster Analysis 86
2.4.1 k-Means 87
2.4.2 Alternatives to k-means 88
2.5 Association Rule Mining 89
2.6 Conclusions 91
Acknowledgments 92
References 92
Chapter 3 Content-based Recommender Systems: State of the Art and Trends 97
3.1 Introduction 98
3.2 Basics of Content-based Recommender Systems 99
3.2.1 A High Level Architecture of Content-based Systems 99
3.2.2 Advantages and Drawbacks of Content-based Filtering 102
3.3 State of the Art of Content-based Recommender Systems 103
3.3.1 Item Representation 104
3.3.1.1 Keyword-based Vector Space Model 105
3.3.1.2 Review of Keyword-based Systems 106
3.3.1.3 Semantic Analysis by using Ontologies 109
3.3.1.4 Semantic Analysis by using Encyclopedic Knowledge Sources 112
3.3.2 Methods for Learning User Profiles 114
3.3.2.1 Probabilistic Methods and Na¨?ve Bayes 115
3.3.2.2 Relevance Feedback and Rocchio’s Algorithm 116
3.3.2.3 Other Methods 117
3.4 Trends and Future Research 118
3.4.1 The Role of User Generated Content in the Recommendation Process 118
3.4.1.1 Social Tagging Recommender Systems 119
3.4.2 Beyond Over-specializion: Serendipity 120
3.5 Conclusions 123
References 124
Chapter 4 A Comprehensive Survey of Neighborhood-based Recommendation Methods 130
4.1 Introduction 130
4.1.1 Formal Definition of the Problem 131
4.1.2 Overview of Recommendation Approaches 133
4.1.2.1 Content-based approaches 133
4.1.2.2 Collaborative filtering approaches 134
4.1.3 Advantages of Neighborhood Approaches 135
4.1.4 Objectives and Outline 136
4.2 Neighborhood-based Recommendation 137
4.2.1 User-based Rating Prediction 138
4.2.2 User-based Classification 139
4.2.3 Regression VS Classification 140
4.2.4 Item-based Recommendation 140
4.2.5 User-based VS Item-based Recommendation 141
4.3 Components of Neighborhood Methods 143
4.3.1 Rating Normalization 144
4.3.1.1 Mean-centering 144
4.3.1.2 Z-score normalization 145
4.3.1.3 Choosing a normalization scheme 146
4.3.2 Similarity Weight Computation 147
4.3.2.1 Correlation-based similarity 147
4.3.2.2 Other similarity measures 148
4.3.2.3 Accounting for significance 150
4.3.2.4 Accounting for variance 151
4.3.3 Neighborhood Selection 152
4.3.3.1 Pre-filtering of neighbors 152
4.3.3.2 Neighbors in the predictions 153
4.4 Advanced Techniques 154
4.4.1 Dimensionality Reduction Methods 155
4.4.1.1 Decomposing the rating matrix 155
4.4.1.2 Decomposing the similarity matrix 157
4.4.2 Graph-based Methods 158
4.4.2.1 Path-based similarity 159
4.4.2.2 Random walk similarity 160
4.5 Conclusion 162
References 163
Chapter 5Advances in Collaborative Filtering 168
5.1 Introduction 168
5.2 Preliminaries 170
5.2.1 Baseline predictors 171
5.2.2 The Netflix data 172
5.2.3 Implicit feedback 173
5.3 Matrix factorization models 174
5.3.1 SVD 174
5.3.2 SVD++ 176
5.3.3 Time-aware factor model 177
5.3.3.1 Time changing baseline predictors 177
5.3.3.2 Time changing factor model 181
5.3.4 Comparison 182
5.3.4.1 Predicting future days 183
5.3.5 Summary 183
5.4 Neighborhood models 184
5.4.1 Similarity measures 185
5.4.2 Similarity-based interpolation 186
5.4.3 Jointly derived interpolation weights 188
5.4.3.1 Formal model 188
5.4.3.2 Computational issues 190
5.4.4 Summary 191
5.5 Enriching neighborhood models 191
5.5.1 A global neighborhood model 192
5.5.1.1 Building the model 192
5.5.1.2 Parameter Estimation 194
5.5.1.3 Comparison of accuracy 195
5.5.2 A factorized neighborhood model 196
5.5.2.1 Factoring item-item relationships 197
5.5.2.2 A user-user model 200
5.5.3 Temporal dynamics at neighborhood models 203
5.5.4 Summary 205
5.6 Between neighborhood and factorization 205
References 207
Chapter 6Developing Constraint-based Recommenders 210
6.1 Introduction 210
6.2 Development of Recommender Knowledge Bases 214
6.3 User Guidance in Recommendation Processes 217
6.4 Calculating Recommendations 226
6.5 Experiences from Projects and Case Studies 228
6.6 Future Research Issues 230
6.7 Summary 235
References 235
Chapter 7Context-Aware Recommender Systems 239
7.1 Introduction and Motivation 240
7.2 Context in Recommender Systems 241
7.2.1 What is Context? 241
7.2.2 Modeling Contextual Information in Recommender Systems 245
7.2.3 Obtaining Contextual Information 250
7.3 Paradigms for Incorporating Context in Recommender Systems 252
7.3.1 Contextual Pre-Filtering 255
7.3.2 Contextual Post-Filtering 259
7.3.3 Contextual Modeling 260
7.3.3.1 Heuristic-Based Approaches 261
7.3.3.2 Model-Based Approaches 262
7.4 Combining Multiple Approaches 265
7.4.1 Case Study of Combining Multiple Pre-Filters: Algorithms 266
7.4.2 Case Study of Combining Multiple Pre-Filters: ExperimentalResults 267
7.5 Additional Issues in Context-Aware Recommender Systems 269
7.6 Conclusions 271
Acknowledgements 272
References 272
Part IIApplications and Evaluation of RSs 276
Chapter 8Evaluating Recommendation Systems 277
8.1 Introduction 278
8.2 Experimental Settings 280
8.2.1 Offline Experiments 281
8.2.1.1 Data sets for offline experiments 281
8.2.1.2 Simulating user behavior 282
8.2.1.3 More complex user modeling 283
8.2.2 User Studies 283
8.2.2.1 Advantages and Disadvantages 284
8.2.2.2 Between vs.Within Subjects 285
8.2.2.3 Variable Counter Balance 286
8.2.2.4 Questionnaires 286
8.2.3 Online Evaluation 286
8.2.4 Drawing Reliable Conclusions 287
8.2.4.1 Confidence and p-values 288
8.2.4.2 Paired Results 288
8.2.4.3 Unpaired Results 289
8.2.4.4 Multiple tests 290
8.2.4.5 Confidence Intervals 291
8.3 Recommendation System Properties 291
8.3.1 User Preference 292
8.3.2 Prediction Accuracy 293
8.3.2.1 Measuring Ratings Prediction Accuracy 293
8.3.2.2 Measuring Usage Prediction 294
8.3.2.3 Ranking Measures 297
8.3.3 Coverage 301
8.3.3.1 Item Space Coverage 302
8.3.3.2 User Space Coverage 302
8.3.3.3 Cold Start 303
8.3.4 Confidence 303
8.3.5 Trust 305
8.3.6 Novelty 305
8.3.7 Serendipity 306
8.3.8 Diversity 308
8.3.9 Utility 309
8.3.10 Risk 310
8.3.11 Robustness 310
8.3.12 Privacy 311
8.3.13 Adaptivity 312
8.3.14 Scalability 313
8.4 Conclusion 313
References 314
Chapter 9 A Recommender System for an IPTV Service Provider: a Real Large-Scale ProductionEnvironment 318
9.1 Introduction 318
9.2 IPTV Architecture 320
9.2.1 IPTV Search Problems 321
9.3 Recommender System Architecture 322
9.3.1 Data Collection 323
9.3.2 Batch and Real-Time Stages 325
9.4 Recommender Algorithms 327
9.4.1 Overview of Recommender Algorithms 327
9.4.2 LSA Content-Based Algorithm 330
9.4.3 Item-based Collaborative Algorithm 333
9.4.4 Dimensionality-Reduction-Based Collaborative Algorithm 335
9.5 Recommender Services 337
9.6 System Evaluation 338
9.6.1 Off-Line Analysis 340
9.6.2 On-line Analysis 344
9.7 Conclusions 348
References 348
Chapter 10How to Get the Recommender Out of the Lab? 351
10.1 Introduction 352
10.2 Designing Real-World Recommender Systems 352
10.3 Understanding the Recommender Environment 353
10.3.1 Application Model 353
10.3.1.1 Understanding the recommender role in the application 354
10.3.1.2 Understanding the influence of the application implementation 357
10.3.2 User Model 358
10.3.2.1 Understanding who the users are 359
10.3.2.2 Understanding users’ motivations, goals and expectations 360
10.3.2.3 Understanding users’ context 361
10.3.3 Data Model 362
10.3.3.1 Understanding the type of available data to describe items 363
10.3.3.2 Understanding the quality / quantity of metadata 364
10.3.3.3 Understanding the properties of the item set 366
10.3.4 A Method for Using Environment Models 367
10.4 Understanding the Recommender Validation Steps in anIterative Design Process 368
10.4.1 Validation of the Algorithms 368
10.4.2 Validation of the Recommendations 369
10.4.2.1 Card Sorting 369
10.4.2.2 Low fidelity prototyping 370
10.4.2.3 Subjective qualitative evaluation 371
10.4.2.4 Diary studies 372
10.5 Use Case: a Semantic News Recommendation System 373
10.5.1 Context: MESH Project 374
10.5.2 Environmental Models in MESH 375
10.5.2.1 Instantiation of the environmental models 375
10.5.2.2 Links between the different models and constraints on design 375
10.5.3 In Practice: Iterative Instantiations of Models 379
10.6 Conclusion 380
References 380
Chapter 11Matching Recommendation Technologies andDomains 384
11.1 Introduction 384
11.2 RelatedWork 385
11.3 Knowledge Sources 385
11.3.1 Recommendation types 387
11.4 Domain 389
11.4.1 Heterogeneity 389
11.4.2 Risk 390
11.4.3 Churn 390
11.4.4 Interaction Style 391
11.4.5 Preference stability 391
11.4.6 Scrutability 392
11.5 Knowledge Sources 392
11.5.1 Social Knowledge 392
11.5.2 Individual 393
11.5.3 Content 394
11.5.3.1 Domain Knowledge 394
11.6 Mapping Domains to Technologies 395
11.6.1 Algorithms 397
11.6.2 Sample Recommendation Domains 398
11.7 Conclusion 399
Acknowledgements 399
References 399
Chapter 12 Recommender Systems in Technology EnhancedLearning 404
12.1 Introduction 405
12.2 Background 406
12.3 RelatedWork 409
12.4 Survey of TEL Recommender Systems 416
12.5 Evaluation of TEL Recommenders 421
12.6 Conclusions and further work 425
Acknowledgements 426
References 426
Part IIIInteracting with Recommender Systems 433
Chapter 13On the Evolution of Critiquing Recommenders 434
13.1 Introduction 434
13.2 The Early Days: Critiquing Systems/Recognised Benefits 435
13.3 Representation & Retrieval Challenges for CritiquingSystems
13.3.1 Approaches to Critique Representation 437
13.3.1.1 Over-critiquing & protracted recommendation dialogues
13.3.1.2 Critique redundancy & hidden feature-dependency
13.3.1.3 Limited product-space vision 439
13.3.1.4 Weak relevance of presented feedback options 441
13.3.1.5 Limitations of domain & preference driven approaches
13.3.1.6 Restricted user control 442
13.3.2 Retrieval Challenges in Critique-Based Recommenders 445
13.3.2.1 Preference inconsistency and longevity 446
13.3.2.2 Diminishing choices & unreachability
13.3.2.3 Refining recommendation retrievals 450
13.3.2.4 Multi-user preference handing 451
13.4 Interfacing Considerations Across Critiquing Platforms 453
13.4.1 Scaling to Alternate Critiquing Platforms 453
13.4.2 Direct Manipulation Interfaces vs Restricted User Control 455
13.4.3 Supporting Explanation, Confidence & Trust
13.4.4 Visualisation, Adaptivity, and Partitioned Dynamicity 458
13.4.5 Respecting Multi-cultural Usability Differences 460
13.5 Evaluating Critiquing: Resources, Methodologies andCriteria 460
13.5.1 Resources & Methodologies
13.5.2 Evaluation Criteria 461
13.6 Conclusion / Open Challenges & Opportunities
References 464
Chapter 14 Creating More Credible and Persuasive Recommender Systems: The Influence of Source Characteristics on Recommender SystemEvaluations 469
14.1 Introduction 469
14.2 Recommender Systems as Social Actors 470
14.3 Source Credibility 471
14.3.1 Trustworthiness 472
14.3.2 Expertise 472
14.3.3 Influences on Source Credibility 472
14.4 Source Characteristics Studied in Human-HumanInteractions 473
14.4.1 Similarity 473
14.4.1.1 Expertise Judgments 473
14.4.1.2 Trustworthiness Judgments 473
14.4.2 Likeability 474
14.4.3 Symbols of Authority 474
14.4.4 Styles of Speech 475
14.4.5 Physical Attractiveness 475
14.4.6 Humor 475
14.5 Source Characteristics in Human-Computer Interactions 476
14.6 Source Characteristics in Human-Recommender SystemInteractions 477
14.6.1 Recommender system type 477
14.6.2 Input characteristics 478
14.6.3 Process characteristics 479
14.6.4 Output characteristics 479
14.6.5 Characteristics of embodied agents 481
14.7 Discussion 482
14.8 Implications 482
14.9 Directions for future research 484
References 485
Chapter 15 Designing and Evaluating Explanations forRecommender Systems 492
15.1 Introduction 492
15.2 Guidelines 494
15.3 Explanations in Expert Systems 494
15.4 Defining Goals 495
15.4.1 Explain How the System Works: Transparency 496
15.4.2 Allow Users to Tell the System it is Wrong: Scrutability 498
15.4.3 Increase Users’ Confidence in the System: Trust 498
15.4.4 Convince Users to Try or Buy: Persuasiveness 500
15.4.5 Help Users Make Good Decisions: Effectiveness 501
15.4.6 Help Users Make Decisions Faster: Efficiency 503
15.4.7 Make the use of the system enjoyable: Satisfaction 504
15.5 Evaluating the Impact of Explanations on theRecommender System 505
15.5.1 Accuracy Metrics 506
15.5.2 Learning Rate 506
15.5.3 Coverage 507
15.5.4 Acceptance 507
15.6 Designing the Presentation and Interaction withRecommendations 508
15.6.1 Presenting Recommendations 508
15.6.2 Interacting with the Recommender System 509
15.7 Explanation Styles 510
15.7.1 Collaborative-Based Style Explanations 513
15.7.2 Content-Based Style Explanation 514
15.7.3 Case-Based Reasoning (CBR) Style Explanations 516
15.7.4 Knowledge and Utility-Based Style Explanations 517
15.7.5 Demographic Style Explanations 518
15.8 Summary and future directions 518
References 520
Chapter 16Usability Guidelines for Product Recommenders Based on Example Critiquing Research 524
16.1 Introduction 525
16.2 Preliminaries 526
16.2.1 Interaction Model 526
16.2.2 Utility-Based Recommenders 528
16.2.3 The Accuracy, Confidence, Effort Framework 530
16.2.4 Organization of this Chapter 531
16.3 RelatedWork 531
16.3.1 Types of Recommenders 531
16.3.2 Rating-based Systems 532
16.3.3 Case-based Systems 532
16.3.4 Utility-based Systems 532
16.3.5 Critiquing-based Systems 533
16.3.6 Other Design Guidelines 533
16.4 Initial Preference Elicitation 534
16.5 Stimulating Preference Expression with Examples 538
16.5.1 How Many Examples to Show 540
16.5.2 What Examples to Show 540
16.6 Preference Revision 543
16.6.1 Preference Conflicts and Partial Satisfaction 544
16.6.2 Tradeoff Assistance 545
16.7 Display Strategies 547
16.7.1 Recommending One Item at a Time 547
16.7.2 Recommending K best Items 548
16.7.3 Explanation Interfaces 549
16.8 A Model for Rationalizing the Guidelines 550
16.9 Conclusion 554
References 554
Chapter 17 Map Based Visualization of Product Catalogs 559
17.1 Introduction 559
17.2 Methods for Map Based Visualization 561
17.2.1 Self-Organizing Maps 562
17.2.2 Treemaps 563
17.2.3 Multidimensional Scaling 565
17.2.4 Nonlinear Principal Components Analysis 565
17.3 Product Catalog Maps 566
17.3.1 Multidimensional Scaling 567
17.3.2 Nonlinear Principal Components Analysis 570
17.4 Determining AttributeWeights using Clickstream Analysis 571
17.4.1 Poisson Regression Model 572
17.4.2 Handling Missing Values 572
17.4.3 Choosing Weights Using Poisson Regression 573
17.4.4 Stepwise Poisson Regression Model 574
17.5 Graphical Shopping Interface 574
17.6 E-Commerce Applications 575
17.6.1 MDS Based Product Catalog Map Using Attribute Weights 576
17.6.2 NL-PCA Based Product Catalog Map 580
17.6.3 Graphical Shopping Interface 582
17.7 Conclusions and Outlook 585
Acknowledgements 586
References 586
Part IVRecommender Systems and Communities 589
Chapter 18 Communities, Collaboration, and RecommenderSystems in PersonalizedWeb Search 590
18.1 Introduction 590
18.2 A Brief History ofWeb Search 592
18.3 The Future ofWeb Search 594
18.3.1 Personalized Web Search 595
18.3.2 Collaborative Information Retrieval 599
18.3.3 Towards Social Search 601
18.4 Case-Study 1 - Community-BasedWeb Search 602
18.4.1 Repetition and Regularity in Search Communities 603
18.4.2 The Collaborative Web Search System 604
18.4.3 Evaluation 607
18.4.4 Discussion 609
18.5 Case-Study 2 - Web Search. Shared. 609
18.5.1 The HeyStaks System 610
18.5.2 The HeyStaks Recomendation Engine 613
18.5.3 Evaluation 615
18.5.4 Discussion 618
18.6 Conclusions 618
Acknowledgements 619
References 620
Chapter 19Social Tagging Recommender Systems 626
19.1 Introduction 627
19.2 Social Tagging Recommenders Systems 628
19.2.1 Folksonomy 629
19.2.2 The Traditional Recommender Systems Paradigm 630
19.2.3 Multi-mode Recommendations 631
19.3 RealWorld Social Tagging Recommender Systems 632
19.3.1 What are the Challenges? 632
19.3.2 BibSonomy as Study Case 633
19.3.2.1 System Description 633
19.3.2.2 Recommendations in BibSonom 633
19.3.2.3 Technological and Infrastructure Requirements 634
19.3.3 Tag Acquisition 635
19.4 Recommendation Algorithms for Social Tagging Systems 637
19.4.1 Collaborative Filtering 637
19.4.2 Recommendation based on Ranking 641
19.4.2.1 Ranking based on Tensor Factorization 641
19.4.2.2 FolkRank 643
19.4.3 Content-Based Social Tagging RS 645
19.4.3.1 Text-Based 645
19.4.3.2 Image-Based 646
19.4.3.3 Audio-Based 647
19.4.4 Evaluation Protocols and Metrics 648
19.5 Comparison of Algorithms 650
19.6 Conclusions and Research Directions 651
References 653
Chapter 20Trust and Recommendations 656
20.1 Introduction 656
20.2 Computational Trust 658
20.2.1 Trust Representation 659
20.2.2 Trust Computation 661
20.2.2.1 Propagation 661
20.2.2.2 Aggregation 665
20.3 Trust-Enhanced Recommender Systems 666
20.3.1 Motivation 667
20.3.2 State of the Art 669
20.3.2.1 Mining a Trust Network 670
20.3.2.2 Automatic Trust Generation 673
20.3.3 Empirical Comparison 675
20.3.3.1 Data Sets 675
20.3.3.2 Coverage 677
20.3.3.3 Accuracy 679
20.3.3.4 Conclusion 680
20.4 Recent Developments and Open Challenges 681
20.5 Conclusions 683
References 683
Chapter 21Group Recommender Systems:Combining Individual Models 687
21.1 Introduction 687
21.2 Usage Scenarios and Classification of GroupRecommenders 689
21.2.1 Interactive Television 689
21.2.2 Ambient Intelligence 689
21.2.3 Scenarios Underlying Related Work 690
21.2.4 A Classification of Group Recommenders 691
21.3 Aggregation Strategies 692
21.3.1 Overview of Aggregation Strategies 692
21.3.2 Aggregation Strategies Used in Related Work 693
21.3.3 Which Strategy Performs Best 695
21.4 Impact of Sequence Order 696
21.5 Modelling Affective State 698
21.5.1 Modelling an Individual’s Satisfaction on its Own 699
21.5.2 Effects of the Group on an Individual’s Satisfaction 700
21.6 Using Affective State inside Aggregation Strategies 701
21.7 Applying Group Recommendation to Individual Users 703
21.7.1 Multiple Criteria 703
21.7.2 Cold-Start Problem 705
21.7.3 Virtual Group Members 707
21.8 Conclusions and Challenges 707
21.8.1 Main Issues Raised 707
21.8.2 Caveat: Group Modelling 708
21.8.3 Challenges 708
Acknowledgments 711
References 711
Part VAdvanced Algorithms 713
Chapter 22 Aggregation of Preferences in RecommenderSystems 714
22.1 Introduction 714
22.2 Types of Aggregation in Recommender Systems 715
22.2.1 Aggregation of Preferences in CF 717
22.2.2 Aggregation of Features in CB and UB Recommendation 717
22.2.3 Profile Construction for CB, UB 718
22.2.4 Item and User Similarity and Neighborhood Formation 718
22.2.5 Connectives in Case-Based Reasoning for RS 720
22.2.6 Weighted Hybrid Systems 720
22.3 Review of Aggregation Functions 721
22.3.1 Definitions and Properties 721
22.3.1.1 Practical Considerations in RS 723
22.3.2 Aggregation Families 725
22.3.2.1 Quasi-Arithmetic Means 725
22.3.2.2 OWA Functions 726
22.3.2.3 Choquet and Sugeno integrals 727
22.3.2.4 T-Norms and T-Conorms 729
22.3.2.5 Nullnorms and Uninorms 730
22.4 Construction of Aggregation Functions 731
22.4.1 Data Collection and Preprocessing 731
22.4.2 Desired Properties, Semantics and Interpretation 733
22.4.3 Complexity and the Understanding of Function Behavior 734
22.4.4 Weight and Parameter Determination 735
22.5 Sophisticated Aggregation Procedures in RecommenderSystems: Tailoring for Specific Applications 735
22.6 Conclusions 740
22.7 Further Reading 741
Acknowledgements 741
References 742
Chapter 23Active Learning in Recommender Systems 744
23.1 Introduction 744
23.1.1 Objectives of Active Learning in Recommender Systems 746
23.1.2 An Illustrative Example 747
23.1.3 Types of Active Learning 748
23.2 Properties of Data Points 749
23.2.1 Other Considerations 750
23.3 Active Learning in Recommender Systems 751
23.3.1 Method Summary Matrix 751
23.4 Active Learning Formulation 751
23.5 Uncertainty-based Active Learning 755
23.5.1 Output Uncertainty 755
23.5.1.1 Active Learning Methods 756
23.5.1.2 Uncertainty Measurement 756
23.5.2 Decision Boundary Uncertainty 757
23.5.3 Model Uncertainty 758
23.5.3.1 Probabilistic Models 758
23.6 Error-based Active Learning 760
23.6.1 Instance-based Methods 761
23.6.1.1 Output Estimates Change (Y-Change) 761
23.6.1.2 Cross Validation-based 762
23.6.2 Model-based 763
23.6.2.1 Parameter Change-based 764
23.6.2.2 Variance-based 764
23.6.2.3 Image Restoration-based 765
23.7 Ensemble-based Active Learning 765
23.7.1 Models-based 765
23.7.2 Candidates-based 766
23.8 Conversation-based Active Learning 769
23.8.1 Case-based Critique 770
23.8.2 Diversity-based 770
23.8.3 Query Editing-based 771
23.9 Computational Considerations 771
23.10 Discussion 772
Acknowledgments 773
References 773
Chapter 24Multi-Criteria Recommender Systems 777
24.1 Introduction 777
24.2 Recommendation as a Multi-Criteria Decision MakingProblem 779
24.2.1 Object of Decision 780
24.2.2 Family of Criteria 781
24.2.3 Global Preference Model 782
24.2.4 Decision Support Process 783
24.3 MCDM Framework for Recommender Systems: LessonsLearned 784
24.4 Multi-Criteria Rating Recommendation 788
24.4.1 Traditional single-rating recommendation problem 789
24.4.2 Extending traditional recommender systems to includemulti-criteria ratings 790
24.5 Survey of Algorithms for Multi-Criteria RatingRecommenders 791
24.5.1 Engaging Multi-Criteria Ratings during Prediction 792
24.5.1.1 Heuristic approaches 792
24.5.1.2 Model-based approaches 795
24.5.2 Engaging Multi-Criteria Ratings during Recommendation 799
24.5.2.1 Related work: multi-criteria optimization 800
24.5.2.2 Designing a total order for item recommendations 800
24.5.2.3 Finding Pareto optimal item recommendations 801
24.5.2.4 Using multi-criteria ratings as recommendation filters 802
24.6 Discussion and FutureWork 803
24.7 Conclusions 805
References 806
Chapter 25 Robust Collaborative Recommendation 812
25.1 Introduction 812
25.2 Defining the Problem 814
25.2.1 An Example Attack 816
25.3 Characterising Attacks 817
25.3.1 Basic Attacks 817
25.3.1.1 Random Attack 817
25.3.1.2 Average Attack 817
25.3.2 Low-knowledge attacks 818
25.3.2.1 Bandwagon Attack 818
25.3.2.2 Segment Attack 819
25.3.3 Nuke Attack Models 819
25.3.3.1 Love/Hate Attack 819
25.3.3.2 Reverse Bandwagon Attack 819
25.3.4 Informed Attack Models 820
25.3.4.1 Popular Attack 820
25.3.4.2 Probe Attack Strategy 821
25.4 Measuring Robustness 821
25.4.1 Evaluation Metrics 822
25.4.2 Push Attacks 823
25.4.3 Nuke Attacks 825
25.4.4 Informed Attacks 826
25.4.5 Attack impact 827
25.5 Attack Detection 827
25.5.1 Evaluation Metrics 828
25.5.1.1 Impact on Recommender and Attack Performance 829
25.5.2 Single Profile Detection 829
25.5.2.1 Unsupervised Detection 830
25.5.2.2 Supervised Detection 830
25.5.3 Group Profile Detection 831
25.5.3.1 Neighbourhood Filtering 831
25.5.3.2 Detecting attacks using Profile Clustering 832
25.5.4 Detection findings 834
25.6 Robust Algorithms 835
25.6.1 Model-based Recomendation 835
25.6.2 Robust Matrix Factorisation (RMF) 836
25.6.3 Other Robust Recommendation Algorithms 837
25.6.4 The Influence Limiter and Trust-based Recommendation 838
25.7 Conclusion 839
Acknowledgements 840
References 840
Index 843