Ring-Based Spatial Transformer Outperforms GWR in Predicting Pedestrian Flow Around Tokyo Stations

A recent study introduces a ring-based Spatial Transformer model aimed at understanding non-linear spatial interactions between building functions and pedestrian movement around railway stations. This research, published on arXiv (2608.14660), establishes concentric ring buffers at 100-meter increments, extending up to 800 meters from 100 randomly chosen stations in Tokyo, with each ring serving as a spatial token. Utilizing Self-Attention, the model learns inter-zone interactions from data without pre-existing structural assumptions. GPS-based walking trip counts act as the target variable, with Geographically Weighted Regression (GWR) as the benchmark. In 30 independent trials, the Spatial Transformer consistently surpassed GWR in predictive accuracy. SHAP analysis indicated that features from mid-to-outer distance zones significantly influence pedestrian flow, while the 0-100m zone had minimal impact. The attention matrix revealed that each distance zone primarily focuses on more distant zones, highlighting the role of non-local interactions in pedestrian flow. This research holds significance for urban planning and transportation modeling, providing a data-driven perspective on how building uses at varying distances from stations contribute to pedestrian movement. Conducted in Tokyo, Japan, the study suggests that transformer architectures effectively capture intricate spatial dependencies in urban settings.