Hierarchical semantic fusion in hyperbolic space for next POI recommendation.
Abstract
Next Point-of-Interest (POI) recommendation aims to suggest POIs that users are likely to visit next based on their recent check-ins. Recent studies suggest that POI-POI transitions and user-POI interactions exhibit hierarchical patterns that can be modeled in hyperbolic space, whose exponential volume growth enables compact and expressive representations. However, existing hyperbolic methods insufficiently explore rich contextual semantics, such as category and region information, and thus fail to leverage these signals when modeling user mobility patterns. Besides, empirical analysis implies that check-in frequencies across regions and categories exhibit hierarchical distribution patterns, while both first-order and high-order POI transitions, along with their associated region and category transitions, also imply hierarchical patterns. We hypothesize that effectively exploiting these hierarchical semantics is essential for capturing user mobility patterns and enabling better recommendations. Thus, we propose H2POI, a unified hyperbolic framework that systematically fuses diverse hierarchical semantics through geometry-aware operations for next POI task. H2POI first performs global perspective fusion by integrating first-order POI-POI transitions and user interactions with contextual region and category semantics, thereby constructing expressive entity representations. Then, it proceeds from a trajectory perspective, incorporating region and category semantics to enhance the modeling of check-ins and to construct two complementary trajectory representations that comprehensively capture transition dependencies. Finally, it further incorporates region and category predictions as auxiliary supervision during training, complementing semantic fusion and strengthening next POI recommendation. Experiments on three real-world datasets show that H2POI consistently outperforms state-of-the-art baselines, demonstrating the benefit of jointly modeling hierarchical semantics for next POI recommendation.