Hierarchical Hyperedge Embedding-Based Representation Learning for Group Recommendation

Group recommendation aims to recommend items a group of users. In this work, we study in particular scenario, namely occasional recommendation, where groups are formed ad hoc and users may just constitute for the first time—that is, historical group-item interaction records highly limited. Most state-of-the-art works have addressed challenge by aggregating members’ personal preferences learn representation. However, representation learning is most complex beyond aggregation or fusion member representation, as be different spaces even orthogonal. addition, learned user not accurate due sparsity users’ data. Moreover, similarity terms common members has been overlooked, which, however, great potential improve learning. focus on addressing aforementioned challenges task, devise hierarchical hyperedge embedding-based recommender, HyperGroup. Specifically, propose leverage user-user interactions alleviate issue user-item interactions, design graph neural network-based network enhance individuals’ from their friends’ preferences, which provides solid foundation groups’ preferences. To exploit (i.e., overlapping relationships among groups) more limited connect all sets (a.k.a. hypergraph), treat task preference embedding hyperedges sets/groups) hypergraph, an inductive method proposed. further group-level modeling, develop joint training strategy both same process. We conduct extensive experiments two real-world datasets, experimental results demonstrate superiority our proposed HyperGroup comparison baselines.