Explainable supervised clustering of urban traffic crash severity using SHAP and Self-Organizing Maps
Urban traffic accidents remain a major public health problem, particularly in densely populated cities. This study proposes a supervised clustering framework to identify interpretable typologies of accidents associated with fatal and serious injuries. Using urban accident data from Barcelona (2017-2019), we first develop a machine learning model to predict injury severity and apply SHAP values to explain feature contributions. We then introduce a supervised clustering strategy that integrates SHAP values with self-organising maps (SOM), to group accidents according to common mechanisms that determine severity, rather than the similarity of raw covariates. The method identifies ten distinct and policy-relevant typologies, capturing both well-known and underexplored high-risk scenarios. The results demonstrate that SHAP-based SOM clustering improves the interpretability of predictive models and provides practical segmentation tools for data-driven urban road safety interventions.
Keywords: Explainable Artificial Intelligence Shapley values Kohonen network Road safety planning