Unraveling the spatial heterogeneity of urban traffic crashes: A GeoXGBoost and point-mapping SHAP framework

In urban traffic risk research, traditional models struggle to simultaneously capture the complex nonlinearity and spatial heterogeneity of risk-driving factors, while often lacking fine-grained spatial interpretability. To address these limitations, this study develops a comprehensive traffic risk analysis framework that integrates high predictive accuracy with micro-scale causal attribution.

Using the central urban area of Liyang as an empirical case, an overlapping grid sampling strategy is employed to overcome the constraints of sample sparsity. A GeoXGBoost-based prediction framework is constructed to internalize spatial heterogeneity within the boosting algorithm, supplemented by a local spatial SHAP interpretation framework using a point-mapping strategy to enable precise risk attribution from global statistics to micro-spatial units.

The GeoXGBoost model achieves a superior goodness-of-fit (OOB-predicted R²) of 0.896, significantly outperforming five benchmark models and effectively eliminating the spatial autocorrelation of prediction residuals. Core risk factors, including shopping services, transport facility services, and population density, exhibit pronounced spatial non-stationarity and location-specific functional differentiation, with their driving effects highly dependent on the local road network context.

A deep interactive coupling mechanism of "place-exposure-function" is identified. Standalone static facilities and population scale constitute only potential hazards; however, when superimposed with high-load dynamic traffic flows or highly mixed functions, they trigger an abrupt, nonlinear surge in local crash probabilities. This framework provides a scientific basis for fine-grained, location-specific spatial interventions in urban traffic safety governance.