LLM-SATPOI: A Semantic-Aligned Large Language Model with Temporal Modeling for Next POI Recommendation

Next point-of-interest (POI) recommendation provides critical support for location-based services by predicting the future location of users based on historical trajectory data. Existing POI recommendation methods based on large language models (LLMs) suffer from semantic misalignment between LLMs and recommender systems, limited temporal modeling, and trajectory data sparsity. To address these issues, we propose a semantic-aligned large language model with temporal modeling for next POI recommendation (LLM-SATPOI). Specifically, independent user and POI ID tokens are first introduced to expand the LLM’s vocabulary and construct atomic semantics for POIs and users. Subsequently, a hierarchical time-slot encoding mechanism is developed to model the associative features of user behavior across time periods of varying granularity. Finally, a trajectory-level contrastive learning framework is employed during the training phase, thereby mitigating data sparsity and enhancing the robustness of the model. Experiments conducted on three real-world datasets (NYC, TKY, and CA) demonstrate that LLM-SATPOI improves Acc@1 by \(5.54 \% \sim 9.50 \%\) , and exhibits superior performance in handling short trajectories.