基于SHAP解释的交通事故严重性集成预测模型  

An Ensemble Prediction Model for Traffic Accident Severity Based on SHAP Interpretation

作　　者：刘嘉琪 马社强 王晟由 LIU Jiaqi;MA Sheqiang;WANG Shengyou(School of Traffic Management People's Public Security University of China,Beijing 100038,China)

机构地区：[1]中国人民公安大学交通管理学院,北京100038

出　　处：《交通工程》2025年第1期31-37,50,共8页Journal of Transportation Engineering

基　　金：国家重点研发计划资助(2023YFB4302702)。

摘　　要：本研究旨在提高事故发生后对设施隐患的排查效率,通过构建一个基于SMOTE优化的交通事故严重程度集成预测模型,探究不同设施与交通事故严重程度的关系。采用开源数据集US Accident对模型进行训练验证。首先,引入SMOTE技术处理数据不平衡问题。然后,采用集成学习方法,以逻辑回归为元学习器,结合Adaboost、LightGBM和逻辑回归作为基学习器,通过加权投票策略提升预测性能。结果显示,模型准确率、召回率和F1分数达到0.781 7,优于其他个体模型。进一步引入SHAP值来解释模型的设施特征贡献度,评估显示,交通信号系统是减轻事故严重性的首要措施;公共交通站点作为高密度交通区域,对事故影响较大;停车场因车辆停靠活动风险较高;而交通减速设施与标志能有效提升道路安全性和驾驶体验。This study aims to improve the efficiency of facility hazard identification following traffic accidents by constructing a SMOTE-optimized ensemble prediction model for traffic accident severity.The research investigates the relationship between various facility-related features and traffic accident severity using the publicly available US Accident dataset for training and validation.Initially,SMOTE was employed to address the class imbalance issue.Subsequently,an ensemble learning approach was implemented,utilizing logistic regression as the meta-learner and integrating Adaboost,LightGBM,and logistic regression as base learners with a weighted voting strategy to enhance predictive performance.The results demonstrate that the model achieved an accuracy,recall,and F1 score of 0.7817,outperforming individual models.Furthermore,SHAP values were applied to interpret the contributions of facility-related features.The analysis reveals that traffic signal systems serve as the primary measure to reduce accident severity,public transportation stations significantly influence accidents due to their high-density traffic,parking lots pose elevated risks associated with vehicle parking activities,and traffic calming facilities and signage contribute to improving road safety and driving experience.

关 键 词：交通管理工程 交通事故 设施便利程度 堆叠技术 可解释机制 

分 类 号：U491.2[交通运输工程—交通运输规划与管理]