Beyond Long Tail POIs: Transition-Centered Generalization for Human Mobility Prediction

Human mobility prediction — forecasting a user's next Point of Interest (POI) from historical trajectories — is critical for recommenders and urban planning. A known bottleneck is the long-tail distribution: POIs with few visit records are hard to predict. But the authors show that even for popular POIs, many predictions fail because the specific source-destination transition (e.g., gym to coffee shop) rarely appears in training data. They formulate this as a compositional generalization problem and propose RECAP (transition reconstruction framework for compositional generalization). RECAP extracts two generalizable signals: multi-hop transitivity from a global transition graph (e.g., if A→B and B→C are common, infer A→C) and revisit evidence from the user's own history (did they take a similar path before). A warm-transition holdout training step prevents the model from simply memorizing frequent transitions, forcing it to rely on these transferable signals.

Experiments on multiple real-world mobility datasets show that RECAP consistently improves next-POI prediction accuracy, with clear gains on tail transitions (the rare source-destination pairs). The approach is model-agnostic and can be plugged into existing predictors. For urban planners and app developers, this means more accurate predictions of where users will go next — even for unfamiliar routes — enabling better recommendation, traffic management, and personalized services. The paper is available on arXiv (2605.05771) and the code is expected to be released.