The Scaling Bottleneck of Human Mobility Modeling

Title: The Scaling Bottleneck of Human Mobility Modeling

pic-The Scaling Bottleneck of Human Mobility Modeling

Abstract

Although scaling—training larger models on larger datasets with more compute—has proven effective for improving performance and generalization across many data modalities, a comparable trend has not yet emerged in human mobility modeling. Existing work in human mobility modeling has largely remained limited to relatively small models (<500M parameters) and modest data volumes (<1B samples), leaving it unclear whether, and how, scaling benefits this domain. This gap has limited progress toward human mobility foundation models for broad societal applications. In this work, we construct a large-scale real-world human mobility dataset comprising trillions of timestamped positioning records from millions of individuals, and take a first step toward systematically studying scaling in human mobility modeling. We first examine whether human mobility models exhibit scaling behavior similar to that observed in other domains. Extensive experiments at data scales up to 1011 samples reveal an unexpected failure of conventional scaling in human mobility modeling. We then investigate the source of this scaling failure and trace it to sample-quality issues inherent in the prevailing place-visitation (PV) modeling paradigm. Finally, we propose spatial-interaction behavior (SIB) modeling, a new paradigm designed to overcome this bottleneck and enable effective scaling in human mobility modeling. We show that SIB-based scaling recovers power-law scaling behavior and yields models that can be directly adapted to diverse downstream tasks. These findings identify both a key obstacle to scaling mobility models and a practical route toward scalable human mobility foundation models. Code and supplemental materials: https://fukamaru.github.io/human-mobility-scaling-bottleneck