Abstract
Abstract
Point-of-interest (POI) recommendation aims to infer users' future visits from sparse and heterogeneous check-in records in location-based social networks (LBSNs). Existing methods often model sequential transitions, contextual factors, or graph relations separately, making it difficult to jointly capture latent user groups, POI functional structures, and high-order co-occurrence patterns. To address this problem, this paper proposes clustering-enhanced hypergraph learning (CEHL) for POI recommendation. CEHL first applies attributed graph clustering in a heterogeneous information network to mine user-group and POI-function priors from both topology and node attributes. It then incorporates these priors into a normalized heterogeneous graph Transformer to learn semantic interactions among users, POIs, time slots, categories, regions of interest (ROIs), and their relations. On this basis, dual homogeneous hypergraphs are constructed on the user and POI sides to model high-order relations among users visiting the same POI and POIs visited by the same user. A joint objective combines recommendation optimization with representation alignment to maintain consistency between heterogeneous-graph and hypergraph representations. Experiments on the Foursquare NYC and TKY datasets show that CEHL improves recommendation accuracy and ranking quality over representative baselines, demonstrating the effectiveness of integrating clustering enhancement with high-order hypergraph modeling.