Answers are grounded in this paper and cite their sources.
In traffic safety analysis, studies have conducted traffic accident severity classification and feature importance analysis based on machine learning models. However, most related studies adopt a single machine learning model, making it difficult to mine the data fully. Meanwhile, when analyzing the causes of accidents, related studies did not take into account the potential heterogeneity of features that affect traffic accidents in intersection and nonintersection scenarios for comparative analysis, and the analysis based on Shapley additive explanation (SHAP) sample values may not accurately reveal the relationship between features and accidents. This study attempted to fill the gap. We constructed a two-layer machine learning model architecture and proposed an algorithm, called synthetic minority oversampling technique-stacking tree-structured Parzen estimator (SMOTE-stacking-TPE) (SST) to effectively classify the severity of traffic accidents occurring at intersections and nonintersections, and then explored the features that influence casualty accidents at intersections and nonintersections. Further, we investigated the nonlinear relationship between the density of different land use types and casualty accidents at intersections and nonintersections based on the accumulated local effect (ALE) model. Model evaluation results show that, compared with mainstream models, the SST model achieved better overall performance in terms of accuracy, recall, F1 score, and area under the receiver operating characteristic curve. This demonstrates that the constructed two-layer machine learning model architecture can effectively combine the advantages of different machine learning models and is a more effective architecture for improving model classification performance. Research results show that the features affecting casualty accidents at intersections and nonintersections have heterogeneity. Targeted policy measures should be formulated for intersection and nonintersection areas, respectively. Research results also show that the density of different land use types has a nonlinear relationship and threshold effect with casualty accidents occurring at intersections and nonintersections.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: