Traffic Conflict Identification and Risk Assessment at Ramp Merging Areas Based on Improved XGBoost and SHAP

Ramp merging areas are high-risk zones for traffic accidents due to the high traffic density and frequent changes in vehicle speed. This study proposes a comprehensive approach that integrates machine learning with feature interpretation to identify and assess the risk of traffic conflicts in these areas. First, an improved eXtreme Gradient Boosting (XGBoost) classification model is developed, incorporating the synthetic minority oversampling technique (SMOTE) and Bayesian hyperparameter optimization to enhance its performance in detecting potential traffic conflicts. Then, SHapley Additive exPlanations (SHAP) analysis is employed to quantify the contribution of various vehicle features to traffic conflict identification. The results demonstrate that the proposed model achieves an accuracy of 92.51%, recall rate of 83.51%, precision of 69.86%, and F1 score of 76.08%, significantly outperforming other classic machine learning methods. SHAP analysis revealed that speed and relative speed are the key features influencing traffic conflict identification, while headway and acceleration are secondary. Relative angle, vehicle length, and vehicle width have a lesser impact. Based on these findings, a feature-weighted scoring method for traffic conflict risk assessment is proposed, categorizing risks into high, medium, and low levels, and directly providing risk level assessments based on vehicle traffic characteristics.