Gmr-Rec: Graph Mutual Regularization Learning for Multi-Domain Recommendation
Shuai Li,Yifan Wang,Yongkang Li,Wei Ju,Jiangke Fan,Shan Gao,Ling Ma,Bing Cheng,Xunliang Cai,Sheng Wang,Ming Zhang
DOI: https://doi.org/10.2139/ssrn.4242756
2022-01-01
SSRN Electronic Journal
Abstract:Multi-domain recommender systems play an increasingly important role in helping users find satisfactory items from different domains. By transferring the knowledge from relevant domains, the sparsity challenge within a single domain can be alleviated in a collective manner. However, existing methods mostly focus on sharing/mapping the features of the same users among different domains to transfer knowledge. In fact, user-item interactions of different domains can be formulated as a bipartite graph, and transferring knowledge through the graph is a more explicit way. Meanwhile, these approaches mainly model the common interests of users but overlook domain-specific interests. In this paper, we propose a novel Graph Mutual Regularization Learning framework (GMR-Rec) for the multi-domain recommendation. The key idea of GMR-Rec is to use graph neural networks to transfer knowledge across different domains and use mutual regularization strategies to extract domain-shared and domain-specific user interests. Specifically, domain-shared and domain-specific graphs are constructed based on the historical user-item interactions, and a parallel graph neural network is employed to extract user interests in corresponding graphs. We then use mutual regularization learning strategies to distinguish domain-specific user interests while keeping common user interests shared among domains.Extensive experiments on three real-world datasets show that GMR-Rec consistently outperforms state-of-the-art methods. We further demonstrate the effectiveness and interpretability of the proposed framework through case studies and visualization.