Deep Spatial Evolution Model: Code and Data

Spatial interaction underlying panel data, such as migration flows that occur between consecutive population snapshots, reflects details of spatial evolution but remains challenging to infer from discrete observations. Traditional panel gravity models and optimization-based methods, while interpretable, are limited by linear assumptions and overlook geographic context. Recent deep learning models focus on cross-sectional data and neglect the dynamic nature of flow in spatial panel scenarios. Our work introduces a Deep Spatial Evolution Network (DSEN) to infer the spatial evolution, i.e., flows, from two spatial distribution snapshots, as a simple but general example in spatial panel cases. DSEN incorporates a cross-event context learner to encode the geographic context before and after an event using parallel Graph Attention Networks (GATs), and a spatial evolution decoder to learn latent evolutionary features through gravity-informed neural networks. We validated DSEN on a device-level mobile positioning dataset collected in the Twin Cities Metropolitan Area (TCMA), Minnesota, U.S., to infer human mobility flows during the Christmas holiday in 2021. DSEN improves inference accuracy up to 15.3% over the best baseline method and maintains robustness under the small sample ratio. Exploring evolutionary features reveals meaningful spatial patterns of holiday relocations, offering novel insights for understanding human mobility in a panel view.