Bayesian Matrix Co-Factorization: Variational Algorithm and Cramér-Rao Bound

Matrix factorization is a popular method for collaborative prediction, where unknown ratings are predicted by user and item factor matrices which are determined to approximate a user-item matrix as their product. Bayesian matrix factorization is preferred over other methods for collaborative filtering, since Bayesian approach alleviates overfitting, integrating out all model parameters using variational inference or sampling methods. However, Bayesian matrix factorization still suffers from the cold-start problem where predictions of ratings for new items or of new users' preferences are required. In this paper we present Bayesian matrix co-factorization as an approach to exploiting side information such as content information and demographic user data, where multiple data matrices are jointly decomposed, i.e., each Bayesian decomposition is coupled by sharing some factor matrices. We derive variational inference algorithm for Bayesian matrix co-factorization. In addition, we compute Bayesian Cramér-Rao bound in the case of Gaussian likelihood, showing that Bayesian matrix co-factorization indeed improves the reconstruction over Bayesian factorization of single data matrix. Numerical experiments demonstrate the useful behavior of Bayesian matrix co-factorization in the case of cold-start problems.