Grid-based Gaussian Processes Factorization Machine for Recommender Systems

Matrix Factorization (MF) is an effective approach to Collaborative Filtering (CF) in the field of recommender systems. However, the traditional Matrix Factorization method can only model the linear interaction of latent features between users and items, which is not convincing in reality. In addition, when only sparse dataset of small size is available and the historic behavior logs of most users are scarce, the traditional Matrix Factorization method behaves badly. In this paper, we propose a new model called Grid-based Gaussian Processes Factorization Machine (GGPFM), which is based on Gaussian Processes (GP), to capture the nonlinear relationship between users and items. The generic inference and learning algorithms for GP regression have cubic complexity with respect to the size of dataset. Rather than using off-the-shelf GP model, we endow the latent features with grid structures in order to decrease the model's complexity. Finally, we empirically show that our approach outperforms the state of the art method Gaussian Processes Factorization Machine (GPFM) and the traditional MF method significantly in cases where the data is sparse and most users are inactive.