Analyzing social influence in recommender systems

Abstract

The aim of this thesis is to formalize and investigate the degree of impact that social connections have to the rating behavior of users, by studying publicly available datasets. This research provides a better understanding of specific aspects of social connections that are important when making recommendations, and thus contributes towards designing more effective social recommenders. It is clear that social recommenders are successful: they improve the prediction accuracy in all cases examined. However, the assumptions they are based on, have not been thoroughly studied. When and how should we use social connections to augment collaborative filtering techniques? Is there room for improvement in existing techniques? To answer such questions, we must first examine in detail the relationships between the two sources of information available to a social recommender, the ratings behavior and the social connections. We note that although previous work has investigated some of these relationships, it has done so in a non-systematic way. Our methodological approach considers the aforementioned two views, the historical rating behavior (V1), and the social connections (V2) of users. The goal of our study is to examine whether connections between these two views exist. More concretely, we define several attributes capturing important aspects of each view, and then observe whether there is a correlation between them. We discern three types of attributes: those that concern users individually, which we call level 1 (L1) attributes; those that quantify relations between pairs of users, which we call level 2 (L2) attributes and level 3 (L3) those that correspond to user communities. For each level, we pose research questions that help us explore and understand the connections between the two views. This thesis finds that there exist significant connections between the rating and the social behavior (the two views) in social recommenders at all levels (among individuals, among pairs and friends, and within communities). It also finds that the strength of the connections depends on the specific attributes examined, and is often directional, rather than bi-directional. Moreover, when looking at individuals in the context of their social circle (the communities they belong to), our analysis shows that although various social recommenders have comparable effectiveness, they differ in the impact they have on individuals preferences. Overall, this thesis makes several concrete contributions towards a better understanding of the impact of social connections to the rating behavior of users in social recommender systems. Furthermore, this work proposes a social recommender that is as effective as existing techniques and treats users rather fairly.