Inferring directional spatial evolution pattern from sequential snapshots of spatial distribution

Abstract

Inferring the spatial evolution pattern, for example, estimating human mobility flows from sequential snapshots of visitation maps, can provide insights into the driving force underlying the human activity dynamics. Accurate inferences will benefit urban management and transportation optimization, especially during social events that cause significant changes in human activities.

Previous studies mostly focused on generating population movements based on a single spatial distribution snapshot, while the evolution process that leads to the change of spatial distribution over time remains understudied. Existing inference models do not consider the directional nature of spatial evolution, resulting in non-directional flows in the inferred spatial interaction patterns.

In this study, we introduce a deep learning-based inference model, Bidirectional Spatial Evolution Network (BiSEN), to retrieve the bidirectional spatial interaction patterns from sequential snapshots of spatial distribution. First, a graph attention module is designed to generate location embeddings that capture the complex roles of locations during evolution, considering the geospatial adjacency and configurations of location characteristics. Second, a simple feed-forward neural network is utilized to model the non-linear relationship between location embeddings and the bidirectional evolutionary process.

Utilizing a large-scale device-level mobile positioning dataset, the proposed BiSEN model was tested on inferring the human travels before the Christmas holiday in the Twin Cities Metropolitan Area (TCMA), Minnesota, U.S. Compared with baseline models, the BiSEN model achieves significant performance improvement in reproducing the bidirectional mobility flows between two human visitation snapshots. Our work sheds light on how spatially-explicit artificial intelligence can be empowered to discover unknown spatial evolution knowledge.