Multi-behavior Enhanced Graph Neural Networks for Social Recommendation

Social recommendation has gained more and more attention by utilizing the social relationships among users, alleviating the data sparsity problem in collaborative filtering. Most existing social recommendation approaches treat the preference propagation process coarse-grained, ignoring the different diffusion patterns targeting corresponding interaction behaviors. However, this may be inappropriate because of the interplay between multi-behavior and social relations. Therefore, in this paper, we propose a novel framework, MB-Soc, for Multi-Behavior Enhanced Social Recommender, to model the mutual effect between users' multiple behaviors and social connections. In MB-Soc, we first devise a single behavior-based social diffusion module to depict behavioral trust propagation. Moreover, to support behavior integration, we propose an intent embedding to ensure behavior independency. In addition, we design a Self-Supervised Learning-based behavior integration module to capture the correlations among multiple behaviors. Extensive experiments conducted on two real-world datasets demonstrate the effectiveness of our model.