An extension of matrix factorization with mixture model and its applications

Abstract

Recommendation systems suggest items to an user, based on history of evaluation which is represented as a matrix with missing values. Matrix factorization is a well-known method for predicting missing values of the matrix and its cost function corresponds to the log-likelihood with Gaussian model, whose mean is assumed to be factorizes. However, in real datasets, a group of users forms a cluster having different characteristics and the conventional model is not appropriate for the situation. To deal with the situation, we propose matrix factorization method with a mixture model and investigate its performance.