Advances in Knowledge Discovery and Data Mining

Similarity in people to people (P2P) recommendation in social networks is not symmetric, where both entities of a relationship are involved in the reciprocal process of determining the success of the relationship. The widely used memory-based collaborative filtering (CF) has advantages of effectiveness and efficiency in traditional item to people recommendation. However, the critical step of computation of similarity between the subjects or objects of recommendation in memory-based CF is typically based on a heuristically symmetric relationship, which may be flawed in P2P recommendation. In this paper, we show that memory-based CF can be significantly improved by using a novel asymmetric model of similarity that considers the probabilities of both positive and negative behaviours, for example, in accepting or rejecting a recommended relationship. We present also a unified model of the fundamental principles of collaborative recommender systems that subsumes both user-based and item-based CF. Our experiments evaluate the proposed approach in P2P recommendation in the real world online dating application, showing significantly improved performance over traditional memory-based methods.