LSR-POI: An LLM-driven Semantic and Retrieval Enhancement Framework for POI Recommendation

Point-of-Interest (POI) recommendation aims to predict users’ potential future visits based on their historical check-in sequences. However, in real-world scenarios, most users have interacted with only a few POIs, while the majority of POIs are visited by only a small number of users. This leads to severe long-tail user and long-tail POI problems. Although several studies have attempted to address these challenges, user behaviors in POI recommendation are typically more personalized and the interaction sequences are much sparser, making it difficult for traditional methods to be directly transferred to POI recommendation scenarios. To tackle these issues, we propose an LLM-driven semantic and retrieval enhancement framework LSR-POI to improve POI recommendation. Specifically, an LLM Semantic Graph Encoder denoises multi-modal POI content and encodes it into structured semantic embeddings, which better represents long-tail POIs. Meanwhile, an LLM Retrieval Augmentor leverages interaction knowledge from semantically similar users to refine preference modeling for long-tail users. Extensive experiments on three real-world datasets demonstrate that LSR-POI consistently outperforms state-of-the-art baselines by 7%–11%, with particularly significant gains for long-tail users and POIs. These improvements stem from the innovative semantic extraction and retrieval enhancement designs of LSR-POI. ¹