LBMF: Log-Bilinear Matrix Factorization for Recommender Systems

Abstract

Collaborative filtering techniques have been successfully applied in recommender systems recently. In order to improve recommendation accuracy for better user experience, the review texts should be exploited due to its rich information about users’ explicit preferences and items’ features, which cannot be fully revealed only by rating scores. In this paper, we propose an effective algorithm called LBMF to explore review texts and rating scores simultaneously. We directly correlate user and item latent dimensions with each word in review texts and ratings in our model, so semantic word vectors can be easily learned and effectively clustered based on rating values. On the other hand, the learned semantic word vectors can justify the rating values, which can promote better learning of user and item latent vectors for rating prediction. The learned latent dimensions by our model can reasonably explain why users rated items the way they did. This revelation can promote better modeling of user profiles and item information, and enable further analysis of user behaviors. Experimental results on several real-world datasets demonstrate the efficiency and effectiveness of LBMF comparing to the state-of-the-art models.