GroupDiff: Exploring A Unified Graph Structure and High-order Interactions for Group Recommendation

With the development of social networks, it is becoming popular and convenient for people to join groups and take activities together. How to recommend to a group of users is a challenging problem, since some groups might be ad-hoc with sparse historical interactions, which results in the cold-start problem. Besides, group members usually have different preferences, making it difficult to reach a consensus. Although prior works have proposed to address such issues, they only take the first-order interactions among groups, users and items into consideration, while the high-order interactions are neglected, which are actually very informative for representation learning of group preferences. Furthermore, the interactions among groups, users and items are complex, making it difficult to model and learn from their high-order relationships. To address these issues, in this paper, we design a unified graph structure with heterogeneous nodes. It integrates all three types of interactions, i.e. user-item, group-user and group-item interactions, in one framework, which can facilitate the information diffusion and utilization along with the high-order interactions, and thus be beneficial to group recommendation mechanism design. Besides, based on the unified graph, we also propose GroupDiff, which is a graph convolution network-based model and explores the high-order relationships for representation learning of groups, users and items. Experiment results show the effectiveness of our proposed model.