Enhancing next POI recommendation via multi-graph modeling and multi-granularity contrastive learning

Abstract

Abstract Next Point-of-Interest (POI) recommendation is a core task in location-aware services and mobile applications, aiming to predict a user’s next likely location based on historical check-in behavior. Although recent studies leveraging sequential modeling and graph neural networks have shown promising results, they still struggle to capture high-order transitions between POIs and to jointly model multiple semantic views. In this paper, we propose MGMCL, a novel model that integrates multi-graph and multi-contrastive learning to enhance sequential modeling for next-POI recommendation. Specifically, a geospatial relation graph and a trajectory transition graph are constructed from user check-in data to model geographic preferences and transition dynamics, respectively. To better capture spatial semantics and behavioral patterns, MGMCL employs location-aware attention and a gated graph neural network. A local perturbation-based contrastive learning strategy is introduced to improve the robustness and structural consistency of POI embeddings, while a global cross-view contrastive objective aligns representations across graphs, facilitating collaborative modeling of multi-source semantics. Finally, a hybrid Transformer-LSTM encoder-decoder architecture jointly captures users’ long-term preferences and short-term behavioral dynamics, enabling more accurate and personalized POI recommendations. Experiments on two real-world datasets show that MGMCL consistently outperforms baseline methods in the next-POI recommendation task.