Exploiting Cluster Structure in Probabilistic Matrix Factorization

Low-rank matrix factorization is a basic model for collaborative filtering. The low-rank matrix approximation model is equivalent to represent users and items by latent factors, and rating is obtained by calculating the inner product of factors. Most low-rank matrix approximation models assume the latent factors come from a common Gaussian distribution. However, users with similar preferences or items of the same type tend to have similar factors, thus there exists cluster structure underlying user factors and item factors. In this paper, we exploit the cluster structure in the low-rank matrix factorization model to improve prediction accuracy. Experimental results on MovieLens-1m and MovieLens-10m datasets demonstrate the effectiveness of the proposed methods.