Investigating the contributing factors of urban crash levels: A novel stacking integrated learning framework

Crash contributing factors identification plays crucial role in preventing crashes and informing decision-making processes. However, current methods heavily rely on subjective judgments by technical experts, neglecting a comprehensive and scientific analysis. To address this gap, we propose a research framework that utilizes stacking integrated learning to predict crash risk levels and identify crash contributing factors. This framework first utilizes ArcGIS to construct adaptive buffers and obtain multi-source data (i. e., street view data, POI data, crash data). Then we integrate four single models (e.g., RF, SVM, XGBoost, NBC) as the base models, and treat their prediction results on the training data as a new feature set. We then train a meta model with the true labels as the supervision signal, thereby fusing the results of models into the meta model. The most effective model is determined through performance comparison, and the mean relative importance (MRI) of crash contributing factors on a spatial scale is derived from the results. The research results indicate that POIs have the highest MRI of 5.99% among all crash contributing factors ranking, followed by road signage markings (1.98%). These results contribute to enhancing transportation system performance, and intervening in hazardous areas in advance to reduce the occurrence of crashes.