Relational and Multi-Temporal Preference Learning for Next-POI Recommendation

Next point-of-interest (POI) recommendation requires an intelligent representation of user mobility that accounts for relational dependencies, evolving preferences, and irregular movement intervals. However, many existing methods focus primarily on transitions between consecutive check-ins. They therefore provide limited representations of broader POI relationships, behavioral similarities among users, and the complementary roles of long- and short-term preferences. This study proposes a relational and multi-temporal preference learning framework for next-POI recommendation. The framework first constructs a POI–POI graph from geographical proximity and interaction correlations and a user–user graph from similarities in visitation behavior. Multi-head graph attention networks are then used to learn relation-aware representations of POIs and users by assigning different importance weights to their neighboring nodes. To model preference dynamics, the framework learns long-term preferences through attentive aggregation of historical POI, category, and temporal features. It simultaneously captures short-term intentions through parallel long short-term memory networks that model location and category sequences while incorporating irregular temporal and spatial intervals. A personalized fusion mechanism subsequently determines the relative contributions of long- and short-term preferences for each user. Experiments on three real-world Yelp datasets show that the proposed framework consistently outperforms the strongest competing baseline across the reported Recall@K, NDCG@K, and MAP@K settings. Ablation experiments further demonstrate the contributions of relational graph representations, irregular-interval information, dual-sequence modeling, and personalized preference fusion.