Geography-Aware Large Language Models for Next POI Recommendation

The next Point-of-Interest (POI) recommendation task, which predicts a user's subsequent destination based on historical mobility data, is a key problem in location-based services and personalized data management. Accurate next POI recommendation requires effective modeling of geographic dependencies and collaborative transition relations between POIs. Although Large Language Models (LLMs) exhibit strong semantic and contextual reasoning capabilities, their direct applications to spatial recommendation remains limited due to two major challenges: (1) Sparse coordinate semantics, where the infrequent appearance of specific GPS coordinates impairs fine-grained spatial representation learning; and (2) Lack of transition priors, where the absence of POI-POI transition knowledge constrains LLMs' ability to infer movement patterns. To address these challenges, we propose Geography-Aware Large Language Model (GA-LLM), a unified framework that injects spatial and transition knowledge into LLMs. The Geographic Coordinate Injection Module (GCIM) transforms GPS coordinates into hierarchical and continuous spectral embeddings, enabling multi-scale geographic understanding. The POI Alignment Module (PAM) integrates collaborative transition relations into the LLM's semantic manifold, allowing it to infer global POI associations and generalize to unseen POIs for users. Extensive experiments on three real-world datasets demonstrate that GA-LLM consistently outperforms state-of-the-arts by up to 24.10% in next-POI recommendation accuracy (e.g., Acc@5 and MRR@5). Source code is available at: https://github.com/hugh2009hugh/GA-LLM.