STELLAR: Large Language Model Enhanced Spatiotemporal Semantic Fusion for Urban Next POI Recommendation

Urban next point-of-interest (POI) recommendation is fundamental to location-based services, yet accurate prediction remains challenging because user trajectories are sparse, semantically incomplete, and tightly constrained by time and geography. Recent graph and Transformer models capture transition regularities, while language model-based recommenders introduce textual knowledge, but these two lines of work remain weakly coupled: semantic priors may be plausible in language space yet infeasible under a user’s actual mobility state, and purely behavioral encoders fail to infer latent intent when observations are scarce. We introduce STELLAR, a unified state transition framework that treats language knowledge as a calibrated prior rather than an independent prediction branch. STELLAR first induces semantic mobility priors from structured trajectory descriptions, then projects them into a shared transition state and learns a feasibility gate conditioned on spatial reachability, temporal rhythm, and both short- and long-term behavioral context. A transition memory subsequently propagates this calibrated state to candidate POIs, allowing semantic evidence to influence prediction only when it is consistent with executable mobility. This design directly addresses semantic-behavioral misalignment and avoids ad hoc late fusion. Across three public check-in benchmarks, STELLAR consistently outperforms both purely behavioral and language-augmented baselines in ranking accuracy, robustness to corrupted trajectories, and cold-start generalization.