Enhanced knowledge transfer for collaborative filtering with multi-source heterogeneous feedbacks

Abstract

Collaborative filtering (CF) is a widely used method in recommender systems due to its simplicity and efficiency. But most existing CF methods suffer from data scarcity, which arises from the situation with only a limited number of interactions between users and items. One solution to address is how to introduce Transfer Learning (TL)-based CF methods with heterogeneous feedbacks to deal with multi-source and heterogeneous data, such as rating vs. clicks or rating vs. purchase. However, in some applications, extremely sparse (i.e., sparsity level ≤ 0.1%) target data could cause under-transfer and negative transfer. To address the above issue, we propose an Enhanced Knowledge Transfer for Collaborative Filtering with Multi-Source Heterogeneous Feedbacks (EKT). Specifically, we first propose a weighted collective matrix tri-factorization framework. The proposed framework constrains the auxiliary data and the target data to share the same latent factors of users and items as well as partial cluster-level user-item rating pattern in order to enhance knowledge transfer and alleviate the under-transfer issue. Then, to alleviate the negative transfer issue, we integrate the graph co-regularization terms into a proposed framework, which contains the neighborhood structure information of users and items. At last, we simultaneously minimize the objective function of EKT, which consists of weighted collective matrix tri-factorization and the graph co-regularization of user and item graphs. Since the EKT framework is a non-convex optimization problem, we use an alternating optimization procedure to solve it and further prove its convergence. The experimental results on two benchmark datasets show that our proposed EKT method performs better than other baseline methods at almost all sparsity levels except for the denser case of 1% on ML10M and Netflix.