A new AI model says the shops right outside a train station matter less to foot traffic than the ones five minutes' walk away.
Researchers built a transformer model that treats concentric rings around a train station, spaced 100 meters apart out to 800 meters, as spatial tokens, then uses self-attention to learn how those rings interact. They tested it on 100 randomly chosen stations in Tokyo, using GPS-tracked walking trips as the real-world measure of foot traffic, and compared results against Geographically Weighted Regression, a standard statistical method in urban planning. Across 30 independent trials, the transformer beat the regression baseline every time. A follow-up analysis using SHAP, a technique for explaining model predictions, then showed which rings actually drove the forecasts.
The results undercut a foundational idea in "compact city" planning: that clustering development right next to transit stations is what drives walking. Here, the ring covering the first 100 meters barely mattered to predictions. Attention patterns showed each ring paying the most attention to zones far from it, not next door, implying pedestrian flow gets shaped by the entire walkable catchment area rather than by whatever sits right outside the station gates.
It's one study, on one country's stations, using one specific model architecture, so treat it as a data point rather than a mandate to relitigate transit-oriented development. Still, it's a useful reminder that machine learning is increasingly being pointed at planning assumptions that were locked in long before anyone had the GPS data to actually test them.