1.3 Com paring collaborative and content-based methods 5 Multiagent Hy brid Recommendation 5.1 Introduction: Twothought systems 5.2 Design 5. 2.1 Selection agent adapt ation 5.2.2 Collection agent adapt ation 5.2.3 Profile splitting 5.2.4 Collection agent types 5.2.5 Including explicit collaboration 5.2.6 Advant ages of the ar chitecture 5. 2.7 Related work 5.3 Implement ation 100 5.3.1 Collection agent s 5.3.2 Central router 5.3.3 Selection agents 5.4 Experiment al metho dology 106 5.4.1 Profile accuracy 5.4.2 Comp ari son of sources 5.4.3 Declared topics 107 5.5 Experiment result s 5.5.1 Profile accuracy 108 5.5.2 omparison of sources 109 3.5. 3 Specializat 5.5.4 Seren dipity 5.5.5 User observations 112 5.6 Summary: Exploiting overl aps between interests
4.3 Comparing collaborative and content-based methods ......... 77 5 Multiagent Hybrid Recommendation 81 5.1 Introduction: Two \thought systems" . . . . . . . . . . . . . . . . . . 83 5.2 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 5.2.1 Selection agent adaptation . . . . . . . . . . . . . . . . . . . . 87 5.2.2 Collection agent adaptation ................... 88 5.2.3 Pro le splitting . . . . . . . . . . . . . . . . . . . . . . . . . . 92 5.2.4 Collection agent types ...................... 94 5.2.5 Including explicit collaboration ................. 95 5.2.6 Advantages of the architecture . . . . . . . . . . . . . . . . . . 97 5.2.7 Related work ........................... 99 5.3 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 5.3.1 Collection agents ......................... 101 5.3.2 Central router . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 5.3.3 Selection agents . . . . . . . . . . . . . . . . . . . . . . . . . . 103 5.4 Experimental methodology . . . . . . . . . . . . . . . . . . . . . . . . 106 5.4.1 Pro le accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . 106 5.4.2 Comparison of sources ...................... 107 5.4.3 Declared topics .......................... 107 5.5 Experiment results ............................ 108 5.5.1 Pro le accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . 108 5.5.2 Comparison of sources ...................... 109 5.5.3 Specialization . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 5.5.4 Serendipity ............................ 111 5.5.5 User observations . . . . . . . . . . . . . . . . . . . . . . . . . 112 5.6 Summary: Exploiting overlaps between interests . . . . . . . . . . . . 113 xi
6 Composition of the Recommendation Set 115 6.1 Exploration vs. exploit ation control 117 6.1.1 Design 117 6.1.2 Implement ation 119 6.1.3 Experiments 6.1.4 Related work 126 6.2The“ Slider” interface 6.2.1 Design 6. 2.2 Implement ation 147 148 6.2.4 Observations 6. 2.5 Related work 154 6.2.6 Discussion 6.3 Summary: Deciding on composition 156 7 Conclusions 159 7.1 Future directions for recommender systems 7. 2 Contributions revisited 161 Bibliography 165
6 Composition of the Recommendation Set 115 6.1 Exploration vs. exploitation control . . . . . . . . . . . . . . . . . . . 117 6.1.1 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 6.1.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 119 6.1.3 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 6.1.4 Related work ........................... 126 6.2 The \Slider" interface .......................... 128 6.2.1 Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 6.2.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 147 6.2.3 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 6.2.4 Observations ........................... 150 6.2.5 Related work ........................... 154 6.2.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 6.3 Summary: Deciding on composition . . . . . . . . . . . . . . . . . . . 156 7 Conclusions 159 7.1 Future directions for recommender systems . . . . . . . . . . . . . . . 160 7.2 Contributions revisited . . . . . . . . . . . . . . . . . . . . . . . . . . 161 Bibliography 165 xii
List of tables 2.1 Measures assuming binary-valued relevance, where a +b+c+d al number of documents in the corpus 2. 2 User activities which provide feedback 2.3 Or din al scale on which users rank document s 3.1 Topics used for simulation corpus 3.2 The highest-weighted wor ds and their weight s from the user profile after the end of the experiment to find music-related pages. These words have been stemmed, e. g. regga was originally reggae 3.3 Com pari son of n dpm values given different numbers of preferences held by users, after 10 steps of gradient de scent direct ly on the test set 5. 1 Document token dat a structure 100 5.2 Top t wenty words and ted weights from the profile of a collect i agent specializing in cooking. Some of the word endings have been removed(e. g“ min ce”," minced”and" mincing” all become“minc”) t of the st 109 d to adiust profile 146 pics used for news article collection X11
List of Tables 2.1 Measures assuming binary-valued relevance, where a + b + c + d = jDj, the total number of documents in the corpus. ............. 32 2.2 User activities which provide feedback. ................. 41 2.3 Ordinal scale on which users rank documents. ............. 41 3.1 Topics used for simulation corpus. ................... 57 3.2 The highest-weighted words and their weights from the user pro le after the end of the experiment to nd music-related pages. These words have been stemmed, e.g. regga was originally reggae.. . . . . 62 3.3 Comparison of ndpm values given di erent numbers of preferences held by users, after 10 steps of gradient descent directly on the test set. . . 68 5.1 Document token data structure ..................... 100 5.2 Top twenty words and associated weights from the pro le of a collection agent specializing in cooking. Some of the word endings have been removed (e.g., \mince",\minced" and \mincing" all become \minc") or altered (e.g., \parsley" becomes \parslei") as part of the stemming process. .................................. 109 6.1 Parameters used to adjust pro les given user actions. ......... 146 6.2 Topics used for news article collection. ................. 149 xiii
6.3 Example set of recommendat ions V
6.3 Example set of recommendations. . . . . . . . . . . . . . . . . . . . . 152 xiv
is t o f 2.1 A simple model for a recommen der system 3.1 Basic system with a single user and a single agent(shown split into a collection and a selection phase) 3. 2 The liRa interface 3.3 First five entries from a sample top-ten list produced by a students 3.4 Results of an experiment where only music-rel ated pages where rated highly 3.5 Com parison of the LIRA system against random and human-selected “cool”) pages 64 3.6 Variation of ndpm value after 95 iterations of the recommender system with differing numbers of words used from each document, for I-pref users. Error bars show 95% confiden ce intervals 3.7 Gradient des cent on test set (test of optimal learning), in each case averaged among all possible users with the same preference structure or 500 randomly generated users, whichever is the lesser 3.8 Comparison of ndpm dist ances using different numbers of steps of the gradient descent algorithm, after different numbers of iterations of the recommender system, for 1-pref users 95% confidence intervals shown. 69
List of Figures 2.1 A simple model for a recommender system. . . . . . . . . . . . . . . . 38 3.1 Basic system with a single user and a single agent (shown split into a collection and a selection phase). .................... 46 3.2 The LIRA interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 3.3 First ve entries from a sample top-ten list produced by a student's program. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 3.4 Results of an experiment where only music-related pages where rated highly. ................................... 61 3.5 Comparison of the LIRA system against random and human-selected (\cool") pages. .............................. 64 3.6 Variation of ndpm value after 95 iterations of the recommender system with di ering numbers of words used from each document, for 1-pref users. Error bars show 95% con dence intervals. . . . . . . . . . . . . 65 3.7 Gradient descent on test set (test of optimal learning), in each case averaged among all possible users with the same preference structure, or 500 randomly generated users, whichever is the lesser. ....... 67 3.8 Comparison of ndpm distances using di erent numbers of steps of the gradient descent algorithm, after di erent numbers of iterations of the recommender system, for 1-pref users. 95% con dence intervals shown. 69 xv