GMR-Rec: Graph mutual regularization learning for multi-domain recommendation

Abstract

Multi-domain recommender systems are becoming increasingly significant, as they can alleviate the sparsity challenge and cold-start problem within a single domain by transferring knowledge from related domains in a collective manner. However, existing methods primarily concentrate on the process of sharing or mapping the features of the same users across different domains to facilitate knowledge transfer. Since the user-item interactions can be naturally formulated as bipartite graphs, transferring knowledge via message passing throughout domains would be a more straightforward approach. Moreover, the existing approaches generally pay more attention to modeling the common interests of users, leaving behind the under-explored domain-specific interests. In this paper, we introduce a novel framework, called GMR-Rec, for the multi-domain recommendation, which explicitly transfers knowledge across various domains. Specifically, both domain-shared and domain-specific graphs are constructed using historical user-item interactions, with the parallel graph neural network employed for each of them. Then, mutual regularization strategies are proposed to distinguish domain-specific user interests while preserving common user interests shared across domains. Experimental results on the four real-world datasets show that our model achieves an average improvement of 1.24%, 2.90%, 5.07% and 3.17% in HR@10, and 3.05%, 4.24%, 6.38% and 3.99% in NDCG@10 compared to the state-of-the-art baseline.