CORE-POI: Consistency-Aware Inference for Next POI Recommendation

Abstract

Next Point-of-Interest (POI) recommendation remains challenging not only due to data sparsity, but more fundamentally due to inference-time inconsistency between users' coarse-grained mobility intent (e.g., preferred spatial regions or activity time) and fine-grained POI-level decisions. While recent methods enhance representation learning by incorporating spatio-temporal signals and auxiliary semantics, their final predictions are still made solely based on POI-level scores, without explicitly enforcing consistency with the inferred spatio-temporal context. As a result, even strong models may produce top-ranked recommendations that contradict their own intent predictions. In this work, we identify inference-time spatio-temporal inconsistency as a critical yet underexplored bottleneck that cannot be effectively resolved by increasing architectural complexity alone. To address this issue, we propose CORE-POI, a consistency-aware inference framework that explicitly decouples fine-grained preference modeling from coarse-grained spatio-temporal feasibility. CORE-POI introduces an uncertainty-guided calibration mechanism that adaptively adjusts POI rankings using hybrid feasibility priors—combining intent-driven spatial constraints for exploration and personalized temporal habits for exploitation—while maintaining the original ordering within candidates sharing the same feasibility status. We instantiate CORE-POI on a spatio-temporal Transformer trained with masked pre-training and hierarchical auxiliary supervision, enabling integration into existing models with minor modifications. Extensive experiments on two real-world datasets demonstrate consistent and substantial improvements, particularly in top-rank accuracy and Mean Reciprocal Rank, highlighting that inference-time consistency plays a more critical role in decision quality than marginal architectural gains.