Improving personalized recommendations using community membership information

Abstract

While early recommender systems have mostly focused on numeric ratings to model their interests, recent research in this area has explored a range of other sources that can provide information about user interests, such as their bookmarks, tags, social links, or reviews. One source of information that has received little attention so far is users’ membership in online communities. Online communities frequently evolve around specific topics. Therefore, user membership in a community could be interpreted as a sign of user interests in the topics of a particular community, and furthermore, could apply to personalized recommendations as a source of information. This paper explores the feasibility and the value of using users’ community membership as a source of personalized recommendations for individual users. The first part of the paper focuses on feasibility. It attempts to assess to what extent the interests of users within the same community are truly similar. The second part focuses on the value of this information to personalized recommendations. It suggests several recommendation approaches that use community membership information. It also assesses the comparative quality of recommendations that are generated by these approaches. In particular, we substantiate our approach with one typical social bookmarking system, CiteULike. The results of our study demonstrate that the interests of members of the same communities are significantly closer than the interests of non-connected users. Moreover, we found that recommendation approaches based on community membership produce recommendations that are as accurate as those produced through a collaborative filtering approach, but with better efficiency. The recommendations are also more complete than those produced by a collaborative filtering approach. In addition, for cold-start users who have insufficient bookmarking information to reliably represent their interests, recommendations based on community membership are the most valuable.