MCGT: Multi-Class Graph Model driven Transformer for next POI recommendation

Location-based social networks (LBSNs) play a critical role in the next point-of-interest (POI) recommendation by analyzing user movement patterns. However, it is difficult to accurately define user behavior, as it requires capturing spatial, temporal, and social dependencies together under data sparsity and dynamic change. Existing methods, such as recurrent neural network or global graph model, often ignore localized behavioral patterns and suffer from noise in irrelevant POI nodes. To address these challenges, we propose MCGT, a Multi-Classification Graph Model-Driven Transformer framework that combines community-driven graphs and hierarchical feature learning. MCGT partitions users into communities with similar spatial–temporal trajectories, constructs localized subgraphs to filter noise, and identifies fine-grained community-specific patterns. The framework employs stacked encoders and multi-head attention to model intra-community features and cross-community interactions, while integrating spatial semantics (POI sequences) and temporal semantics (check-in timestamps). Experiments on two real-world datasets (Foursquare-NYC and Foursquare-TKY) confirm MCGT’s superiority, which surpasses state-of-the-art method MCLP with 5.29% and 4.99% gains in ACC@1 and MRR (on Foursquare-TKY), respectively. These results highlight its capability to understand complex user behaviors for precise POI recommendations.