Improving Recommendations with Collaborative Factors

Abstract

Collaborative filtering has become the most popular technique in the field of recommender system to deal with the information overload problem. Most collaborative filtering approaches either based on the intuitive nearest neighbor methods or the scalable latent factor models. In order to benefit from the advantages of these two paradigms, some hybrid strategies are proposed by taking weighted averages on near neighbors’ ratings as effects, or factorizing neighborhood to model interactions and relationships directly. However, these methods usually assume that the latent factors of users/items are independent of each other. Yet in fact, there are relationships among latent factors would affect the performance of recommendations. Motivated by this, in this paper, we introduce the collaborative factors, which are smoothed by near neighbors’ factors, to better capture the intrinsic features for users and items. We further propose a novel collaborative matrix factorization (CoMF) model in order to elaborately incorporate these collaborative factors into latent factor models. Finally, experimental results on two datasets show that our CoMF significantly outperforms some state-of-the-art methods in prediction accuracy.