Abstract

In this paper we propose the rating Correlated Topic Model for rating-based collaborative filtering, which improves the performance of the state-of-the-art Latent Semantic Models in two aspects: (1), making the prediction accuracy more robust to the topic number K; (2), improving the recommendation quality for users with few existed ratings. We achieve our goals by employing the Logistic Normal distribution to capture the correlation between latent topics following the Correlated Topic Model, as well as modifying the generative process to meet the requirement of rating-based collaborative filtering. We derive a parameter estimation algorithm based on variational inference for the proposal. Experiment results on the Movielens data set demonstrate our model's advantages on both referred aspects.

Paper Details