Is the Last Check-In All You Need? Next POI Recommendation: Recall and Rerank

Next Point-of-Interest (POI) recommendation aims to predict the next POI a user will visit based on the historical check-ins, but widely used datasets suffer from severe sparsity, making long-range transition modeling difficult to learn reliably. In this work, we revisit next-POI prediction from a minimalist perspective and point out an extreme short-horizon dominance in these datasets: the most predictive signal often collapses to the last check-in, while higher-order transition patterns quickly become too sparse to exploit. Based on this, we propose K1-POI, a concise recall-and-rerank framework that formulates sparse next-POI prediction as candidate generation and calibration. In the recall stage, we design a window-input self-attention backbone that uses only the most recent check-in (k=1) to retrieve a high-quality top-K candidate set; we further introduce a lightweight contrastive objective to align the backbone representation with the embedding of the ground-truth next POI within the candidate set. In the rerank stage, we propose an efficient Spatio-Temporal Prior Reranker (ST-Reranker) that calibrates the ordering within the top-K candidates using additive, model-agnostic spatio-temporal priors computed from the training split. Experiments on three public datasets show that K1-POI achieves strong performance across all datasets. Beyond performance, we propose a set of simple statistics-based ''guess'' baselines and conduct extensive analyses, showing that benchmark performance is largely explained by user preference, one-step transition routines, and spatio-temporal habits. These findings suggest that mining long-range check-in transition patterns may be a particularly challenging direction under current sparse benchmark settings.