Unveiling spatiotemporal mechanisms of urban traffic: Multi-scale determinants and explainable street-level dynamics of a graph neural network in Berlin

Abstract

Street-level average traffic speeds provide essential information for routing, commuting time estimation, and many transport and environmental assessments. However, publicly available traffic speed data often covers only a small share of the urban road networks. One approach to increase traffic speed data availability is the use of models. Here, we tested a two-stage prediction-interpretation framework that combines a graph neural network (GraphSAGE) with a Geographical and Temporal Weighted Regression (GTWR) approach. While the graph neural network was used for the prediction of traffic speed for our case study region, Berlin, Germany, we used the GTWR approach to provide a post-hoc statistical interpretation of the prediction. We trained the GraphSAGE regression model with Uber Movement speeds as the response variable. OSM-based road attributes, population density, and street-view-derived activity indicators were used as predictors for roughly 72% of the road network without Uber Movement speed data. The model achieved a mean average error for the test data of 5.73 km/h and was capable of reproducing the general spatial pattern of traffic speed across Berlin. GTWR was then applied to estimate local coefficients for the same feature set. GTWR indicated road type and speed limit were positively associated with GraphSAGE-predicted speeds, whereas population density, street-view-derived pedestrian counts, and traffic signal counts were generally associated with lower predicted speeds. Clustering the coefficient vectors revealed four coherent GTWR behavior clusters across the city. These clusters summarize how the model redistributes feature importance across spatial contexts and could help transfer the GraphSAGE model to other cities.