STMamba-GC: Spatiotemporal Mamba with graph contrastive learning for next POI recommendation

Next Point-of-Interest (POI) recommendation serves as a fundamental task in location-based services, enhancing personalized user engagement and supporting intelligent business decisions. Despite advances in sequential modeling and graph neural networks, two key challenges persist. First, sparse user check-in data and dynamic preference shifts hinder the learning of stable and accurate user representations. Second, conventional sequential models employ static fusion mechanisms for spatiotemporal dependencies, limiting their capacity to adapt to varying mobility contexts where temporal and spatial factors contribute differently. To address these challenges, we propose the Spatiotemporal Mamba with Graph Contrastive learning network (STMamba-GC). Our framework constructs three complementary graph views-user transitions, semantic patterns, and geographic proximity-and employs graph augmentation with contrastive learning to extract robust preference representations from sparse data. We further design a spatiotemporal gated fusion Mamba module that dynamically balances temporal patterns and spatial constraints through hierarchical adaptive gating and distance-weighted historical state updates. A multi-task learning mechanism jointly optimizes POI prediction and graph contrastive objectives, providing complementary supervision for improved accuracy. Experiments on NYC and TKY datasets demonstrate that STMamba-GC achieves substantial improvements over state-of-the-art baselines, with relative gains of 10.065% in top-10 accuracy and 8.663% in top-20 accuracy on NYC, and 9.62% in MRR on TKY.