Exploiting Behavior Multiplicity for Social Recommendation Via Graph Neural Networks

In recent years, researchers have attempted to use multi-behavior interaction information and social information to alleviate the data sparsity problem in collaborative filtering. However, existing research studies social recommendation and multi-behavior recommendation separately, making them insufficient in capturing heterogeneous coordination signals from both users and items due to the neglect of cross-user, cross-item interdependent knowledge. In this work, we propose Multi-Behavior Social Recommendation based Graph Neural Networks (MBSR-GNN), which incorporates social information and item knowledge to learn user preferences and item representations. We design a multi-layer cross-domain heterogeneous graph neural network to aggregate feature embeddings of users and items from different types of behavioral interactions, social connections, and item dependencies. Extensive experiments on three real-world datasets demonstrate that MBSR-GNN achieves consistent improvements over multiple state-of-the-arts. Specifically, compared to the best baseline, it achieves a 12.23% improvement in HR@10 and a 22.34% improvement in NDCG@10.