Application of the dual-model interpretability framework of XGBoost-SHAP and GAT-GNNExplainer to investigate the impact of the built environment on traffic accidents at intersections.

Exploring the link of characteristics of the built environment and traffic accidents is crucial for enhancing road safety. Prior research has predominantly utilized the "5D" framework to examine how the built environment influences accident rates. However, intersections, which are high-risk areas for traffic accidents, are significantly influenced by their surrounding streetscape features. These features have not been systematically integrated into the existing "5D" framework of the built environment. Moreover, previous studies considered intersections as isolated units, employing analytical models that presuppose sample independence and neglect the spatial spillover effects of explanatory variables due to the interconnected nature of these units. To overcome these limitations, this study introduces a sixth dimension to the built environment, the Detailing of Intersection Interface, by incorporating surrounding streetscape features at intersections. This dimension is quantified using semantic segmentation techniques to calculate the pixel proportions of static streetscape elements from Street View Images (SVIs). Furthermore, to capture the spatial spillover effects of built environment characteristics at neighboring intersections, a spatial topology graph was constructed, defining intersections as nodes and road connections as edges. A Graph Attention Network (GAT) model was subsequently developed to predict accident occurrences at intersections, effectively accounting for the spatial dependencies of accident risks. A dual-model interpretability framework, combining eXtreme Gradient Boosting-Shapley Additive exPlanations (XGBoost-SHAP) and GAT-GNNExplainer, was established to explore the association between built environment characteristics and accident occurrence from global, neighborhood, and individual perspectives. The results demonstrate that the two-layer GAT model (GAT-2) achieved an accuracy of 80.05%, marking a significant improvement in predictive performance by accounting for the spatial spillover effects of built environment features at nearby intersections. From the global perspective, critical factors influencing intersection traffic accidents include POI density, building density, traffic facility density, and distance to bus. These factors exhibit significant nonlinear and threshold effects, as well as strong interaction effects among them. From the neighborhood perspective, intersection traffic accidents are influenced by local built environment features but also by those of adjacent intersections, and the GAT-GNNExplainer framework effectively identified the key neighboring intersections and their key features contributing to accident occurrence, thereby effectively quantifying the spatial spillover effects of built environment features. From the individual perspective, the XGBoost-SHAP interpretability framework quantified the contributions of local built environment features to the prediction results. Overall, the proposed dual-model interpretability framework not only provides excellent predictive accuracy but also reveals, from a neighborhood perspective, the deeper relationships between the built environment and intersection traffic accidents. These insights offer scientific guidance for developing collaborative strategies to optimize the built environment and enhance intersection safety.