By now, we have many definitions, metrics, and methods for influence, mostly developed by Sociologists(Wasserman, Faust, and Iacobucci 1994). Computer Scientists have also come forward in providing quite a few new metrics( Brin and Page 1998; Kleinberg 1999; Domingos and Richardson 2001; White and Smyth 2003: H, D, and A 2000; Kempe, Kleinberg, and Tardos 2003). Amidst such a mature status of the research on influence the question thus is: what additional research questions remain? We believe that in contrast to some of the other social networks, recommender systems provide us with a unique opportunity to study influence in that recommender systems capture interactions that are formal, quantitative, and observed. The social network can be analyzed directly through data already captured in the computer system. Unlike some of the other social networks, the degree of associations between users is known.( Domingos and Richardson 2001)write: "The key advantage of a collaborative filtering database as a source for minir a social network for viral marketing is that the mechanism by which individuals nfluence each other is known and well understood: it is the collaborative filte algorithm itself. User i influences user j when j sees a recommendation that is partly the result of i's rating. Assuming i and j do not know each other in real life(which, given that they can be anywhere in the world, is likely to be true), there is no other way they can substantially infuence each other. Note that identifying the influential people can bring twin advantages to those whe study group dynamics: (1) The influential people can be directly studied, yielding insight since their choices may be predictive of group choices; or(2)The influential people may be influenced to change the behavior of the group. This potential of changing group behavior can be directed towards either applications that users will appreciate--such as marke with the goal of customer satisfaction, or towards exploitative goals-such as promoting low-quality products. In this thesis, we attempt to explore some of these facets of influence Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Item infuence Items in a recommender system can have varying degrees of influence as well. If we define item rating frequency as the item influence, we may find that some items are rated more often than the others. For example, in MOVIELENS, a movie recommender system, only 10% of the movies have received 70% of all ratings in the system. If we consider how frequently an item is being discussed as its level of influence, we may find hat some items are discussed more often than the others. In the research paper domain where the research papers can be considered as items, a paper may be infuential if it cited often and perhaps cited by many other influential papers. Some items may define important"bridges"between polarized groups. For example, as shown in figure 1. 1,(Krebs 2004) generated a graph of political books purchased from amazon. com around the year 2004, where the relationships between the books indicate if they were purchased by the same person. The graph reveals two clearly distinct groups of books reflecting the ideologies of the two large political parties in the US. Between the two almost distinct groups are a few books that form bridges--without these books, the graph would turn into two isolated subgroups. Therefore the books that form the bridges play an important role and thus can be regarded as influential One point to note from the discussion of the preceding paragraph is that influence of items can be defined in multiple ways. In this thesis, however, we define influence of items by considering a particular goal, that is to improve recommendation accuracy. However,we use item influence information in various applications including effectively learning new user preferences and motivating more evaluations from the users. In the following we explain these aspects of item influence we study in this thesis. In a recommender system, relationships between users are formed by what they rate in common and the way they rate. In other words, a meaningful relationship can be established only if a group of users rate from a common pool of items representing the idiosyncrasy of their tastes. Therefore, we believe that there is a set of items that can bring relatively more Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Esam Evil Party Ho More Triangle ORally Tha way Things oug to Be 用 Mch Mbneywar Daddy?? Enough? PafsctlyLeyal Figure 1. 1: Showing the importance of the"bridge"nodes-without them the graph would degen- erate into two separate cohesive clusters. The graph is drawn using the amazon. com purchase data of political books around 2004, and is extracted from(Krebs 2004) lue to (Harper et al. 2005) show that users value quality recommendations and they are typ ically willing to work for them. Therefore, it is important for a recommender system to guide its members to express their preferences properly so that the profiles they build can help the system in generating accurate recommendations. In this regard, we designate an item as influential if a user's evaluation of it can express her preferences well. In this thesis therefore, we define item influence measures with the purpose of learning user profiles and evaluate them both empirically, by various offline and online experiments, and analytically, by analyzing the inherent properties of the measures Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
We seek the opportunity to use item influence as a way to solve a common problem found in online member-based communities, namely the under-contribution problem Under provided evaluations can be a big problem for recommender systems since the recommen- dations may become unreliable or even impossible to generate. The worst cas the so called free riding ( Sweeny 1973; Andreoni 1988) problem, where users only use oth ers'evaluations, but do not provide evaluations of their own. In this thesis we investigate developing interfaces to convey the information of items influence using a language of value to encourage people to rate more movies 1.3 Contributions In this section we compile a list of high level contributions this dissertati individual contribution is elaborated in the corresponding chapter of this dissertation Chapter 2: Experimental Platform We propose and explain expected utility(rashid et al. 2006a), a new decision theoretic measure to compute recommendation accuracies Chapter 3: The Idea of Users'Influence In Recommender Systems We outline a set of desirable properties of infuence metrics By showing a way to deduce a social network from a CF-based recommender system, we demonstrate how existing influence measures can be adapted to the problem domain Each of the influence measures considered are examined to see how they fall short of meeting the principles of influence. We propose a novel infer asure based on the t(Loo) techniqu that attempts to comply with most of the principles of a good influence measure Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
We provide an algorithm to directly and inexpensively compute LOo-based user influ ence measures on the USER-BASED ANN CF algorithm Chapter 4: ENIPD: An Algorithm-Independent Measure of Influence We propose ENIPD(Rashid, Karypis, and Riedl 2005), a generic user influence measure based on the LOo-technique We provide a list of inexpensive -to-compute factors that may affect ENIPD, and demon strate how to use them to learn effective regression models for ENIPD We build models on two Cf algorithms that have distinct characteristics in order to demonstrate the efficacy of the regression models. We show how ENIPD is related the individual factors, and how the relationships change depending on the underlying CF algorithms being used We empirically demonstrate that user influence ini recommender systems is dependent on the underlying CF algorithm being used Although ENiPD is generic and its definition is not driven by a particular application, re show that it can be applied to particular applications Chapter 5: ENSI: An Influence Measure To Find Early Evaluators We elaborate the aspects of the early evaluation problem in recommender systems first reported by(Avery and Zeckhauser 1997) We propose ENSI, another LOO-based user influence measure with the goal of finding reliable early evaluators We propose an algorithm to find early evaluators for a brand new item in a recom- mender system using ENSL. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission