Weighted nonnegative matrix co-tri-factorization for collaborative prediction

Abstract

Collaborative prediction refers to the task of predicting user preferences on the basis of ratings by other users. Collaborative prediction suffers from the cold start problem where predictions of ratings for new items or predictions of new users' preferences are required. Various methods have been developed to overcome this limitation, exploiting side information such as content information and demographic user data. In this paper we present a matrix factorization method for incorporating side information into collaborative prediction. We develop Weighted Nonnegative Matrix Co-Tri-Factorization (WNMCTF) where we jointly minimize weighted residuals, each of which involves a nonnegative 3-factor decomposition of target or side information matrix. Numerical experiments on MovieLens data confirm the useful behavior of WNMCTF when operating from a cold start. © 2009 Springer-Verlag Berlin Heidelberg.