A Nonlinear Cross-Site Transfer Learning Approach for Recommender Systems

Abstract

This paper targets at utilizing cross-site ratings to alleviate the data sparse problem for recommender systems. The key issue is how to bridge user features between the targeted site and the auxiliary site. In traditional transfer learning models, a linear mapping function is assumed to map the user feature in the auxiliary site into the targeted site. The limitation lies in that when the real data does not follow the linear property, such models will fail to work. Therefore, the motivation of this paper is to identify whether the rating prediction performance in recommender systems can be improved by considering the nonlinear transformation. As a primary study, we propose a nonlinear transfer learning model, and utilize the radial basis function (RBF) kernel to map user features of multiple sites. Through empirical analysis in a real-world cross-site dataset, we demonstrate that by utilizing the nonlinear mapping function RBF kernel, the rating prediction performance is consistently better than previous transfer learning models at a significant scale. It indicates that the nonlinear property does exist in real recommender systems, which has been ignored previously.