List of Figures 1.1“ Bridge” nodes as influential nodes. 2. 1 A part of the MOVIELENS interface 3. 1 Formation of a social network from a CF algorithm 3.2 Inexpensive computation of the Loo-based measures on the USER-BASED 122 kNN CF algorithm to compute a loo-based influence measure are rated matters along with how many are rated for ENIPD 4.3 Comparing entropy and standard deviation of movies 4. 4 Per odels of ENiD 4.5 Effects of the qualitative fac 4.6 Distribution of enid that influent 4.9 Improving coverage with influencers 1 Ar rly ratings can be non-re 5.2 She she steps we took to prepare data nd to inves- tigate the extent an approach is able to handle the early evaluation problem. 75 5.3 Approaches to tackle the early evaluation problem are compared 6.1 Distribution of ENSI on a) USER-BASED kNN and b) ITEM-BASED knn CF 6.2 Nature of ite 6. 3 Limitations of using a set of static neighbors Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
6.4 Item-influence approaches are compared on the USER-BASED kNN CF algorithm. 106 6.5 Item-influence approaches are compared on the ITEM-BASED knn CF algorithm. 107 7.1In stems it is not necessary to be familiar with an item to be able to 7.2 Miced initiative preference elicitation techniques may be necessary if the retail site has many different types of items 119 7. 3 New user signup simulation procedure 7.4 Showing how familiar the movies are to the users as the movies are "presented in batches according to each of the item influence measures we study here .. 124 7.5 Recommendation accuracy results from the offine simulations using various item influence measures on the UsER-BASED kNN CF algorithm 7.6 Recommendation accuracy results from the offine simulations using various item influence measures on the ITEM-BASED knn CF algorithm 126 7.7 Average survey responses 8. 1 The value legend demonstrating how much value an orientation of smiley-faces indicates 8.2 A part of a sample screenshot from a user in the Similar Group experimental 8.3 The Sun Simi itern- influence measure is compared against other measures on the ITEM-BASEd knn Cf algorithm 46 8.4 User actions may not be a reflection of their self reports 48 8.5 Interaction plots between the number of smilies on a movie and experimental conditions 150 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Part Prologue Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Chapter 1 Introduction The Web they sag, is leaving the era of search and entering one of discovery Jeffrey M. O'Brien, Fortune magazine writer Contents 1.1 The Problem Domain: Recommender Systems &e Collaborative Filtering 1.2 Recommender Systems and Influence 1.3 Contributions 1.4 Thesis Roadmap This thesis is about influence in recommender systems. Influence of an entity can be defined as its ability to affect the conduct, behavior, or actions of other entities. Recom mender systems help people find the things they care about from an unmanageably large number of choices by matching up like-minded people. If both of the subjects of influence and recommender systems are familiar to the reader, studying influence in the context of recommender systems might seem a bit out of place at first. The study of influence typically Faust, and Iacobucci 1994), which as the first thing in our mind when we think of recommender systems. Further, most of Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
the research in recommender systems so far has involved ways to improve recommendation accuracy(Adomavicius and Tuzhilin 2005) However, we provide evidence in this thesis that influence is important in recommender systems. We show how we can view relationships between items and people in a recommender system as a social network. It is natural, then, to ask about what the most influential items and users in the system are. Moreover, the knowledge of how to find infuential items and users can enable the system to deliver quality recommendations as the system applies the influence information in various ways, including a) by guiding its members throug the right set of items to evaluate so that the system learns effective user profiles, and b) by selecting reliable users for early evaluations of new items. Influence can be a powerful tool for understanding the workings of recommender systems as well. By investigating how influence forms in a recommender system, we may find strengths and vulnerabilities of the underlying algorithm Broadly speaking, this thesis explores the following: a)the nature of influence in rec ommender systems, b) prior research on influence in other domains and the viability of applying that research to the recommender systems domain, c) new measures of influence based on prior research, extended appropriately for recommender systems, d)the feasibility and implications of meaningful applications of influence. 1 The Problem Domain: Recommender Systems& Collaborative Filtering The advent of the internet has given us far more information than we can handle. For , as of this writing, the blog search engine Technorati tracks more than 63 million blogs, the volunteer-maintained free online encyclopedia wikipedia2 contains more than 1.5 million articles, the reference linking service CrossRef reports to have more than 20 million http://ww Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission