Enhancing urban transportation flow modeling through a graph neural network-based spatially weighted interaction model: a case study of Chicago taxi data
Résumé
Transportation flow modeling is critical in predicting and analyzing movements within networks, enabling effective travel forecasting, infrastructure optimization, and urban planning enhancement. Conventional studies often focus on origins and destinations to understand movement patterns but overlook the spatial characteristics of flows. The Spatially Weighted Interaction Model (SWIM) directly models flows but simplifies them as direct lines between origins and destinations, ignoring their spatial distributions. This limitation hinders SWIM’s ability to address spatial non-stationarity, which reduces prediction accuracy. This study aims to enhance transportation flow modeling by integrating a network-constrained approach into SWIM. It introduces the Graph Neural Network-Based Spatially Weighted Interaction Model (GNN-SWIM). It leverages GNNs to capture complex, nonlinear spatial relationships in transportation flows and estimate spatial similarities more effectively. By incorporating these spatial similarities as weights into SWIM, GNN-SWIM better represents spatial non-stationarity, leading to improved prediction accuracy. A real-world experiment using Chicago taxi trip data demonstrates that GNN-SWIM effectively captures spatial non-stationarity. It estimates coefficients that vary across geographical areas and outperforms conventional methods in prediction accuracy. These findings provide valuable insights for urban planners and policymakers to optimize transportation networks and enhance urban mobility.