Quantifying Geographic Domain Shift for Mobility Model Transferability

Understanding human mobility is crucial for urban planning and policy, and models that generate these mobility flows are increasingly important. A key challenge is their geospatial transferability—how well a model trained in one region performs in another. This research systematically investigates this transferability by introducing the concept of "geographic domain shift," which describes the inherent differences in geographic feature distributions and spatial structures between source and target regions. The study evaluated four representative human mobility generation models using a large dataset of commuting flows across 2,265 US counties. To quantify geographic domain shift, two novel metrics were proposed: mutual information and spatial shift. These metrics were then analyzed using linear mixed-effects regression to determine their association with model transferability. The results highlight substantial spatial heterogeneity and asymmetry in model performance across different regions. Crucially, both the information shift and spatial shift metrics demonstrated statistically significant and complementary power in explaining transferability. This indicates that a model's ability to generalize to new locations depends not only on its design but also on the intrinsic geographic differences between regions. These findings provide a new framework for evaluating and improving the robustness and fairness of human mobility data synthesis across diverse geographic areas, offering valuable insights for GeoAI model development.