SoLGR: Social Enhancement Group Recommendation Via Light Graph Convolution Networks

With the rapid development of social networks, online and offline group activities are becoming more common and diverse. Considering the different interests of group members, the recommendation service for a group is more challenging than the common personalized recommendation. In essence, group recommendation interaction data is a typical heterogeneous graph structure. Therefore, the two challenges to this research are 1) how to learn the representation of groups, users, and items from interaction graphs and social relationship graphs, and 2) how to aggregate the representation of groups, users, and items from these graphs. In this research, we proposed a novel end-to-end Social enhancement Group Recommendation via Light graph convolution networks (SoLGR) to address those challenges. Specifically, we first utilize the meta-path to explore the potential social relationship from the user’s perspective. Afterward, SoLGR deploys a light graph convolution operation on interactive graphs and metapath-based graphs to aggregate the embedding of groups, users, and items on each graph. Finally, the representations of different layers are accumulated and used to achieve predictions for both groups and users. The experimental results show that our proposed model significantly improves the group recommendation performance on two real-world datasets.