Aligned Intrinsic User Factors Knowledge Transfer for Cross-domain Recommender Systems

Abstract

Matrix factorization model of collaborative filtering has been proven to be an effective approach to provide personalized recommendations to users. However, It does not guarantee to obtain high prediction accuracy due to the availability of sparse user-item matrix. One of the leading solutions to overcome this problem is knowledge transfer from other related source domain. Previous work only uses the concept of pooling, i.e., all available domains' training data into a single one, and apply traditional techniques of recommender systems. In this paper, we propose a novel method for cross-domain recommender systems by adapting the learned knowledge, in terms of aligned intrinsic user factors, from existing source domain. We apply matrix factorization to estimate aligned intrinsic user factors and item factors of both source and target domains independently. Thereafter we calculate permutation matrix to align intrinsic user factors of source domain to target domain. At lest, prediction is estimated by multiplying of corresponding intrinsic user factors of source domain, permutation matrix and item factors of target domain. Evaluating the effectiveness of the proposed method, several experiments are done on Amazon co-purchasing dataset. The obtained results demonstrate the improvements of proposed method over other state-of-the-art methods.