Abstract: In order to study the significant factors affecting the severity of four-wheeled vehicle traffic accidents, and to grasp their occurrence characteristics and regularities, 2 966 four-wheeled vehicle accidents in the China in-depth acciddent study (CIDAS) database are modeled to include the 19 factors including people-vehicle-road- environment are used as input and the accident result is used as output. In the research, an evaluation system including accuracy, precision, recall, F-1 score are introduced, and the XGBoost model is compared with LightGBM, random forest and CatBoost models side by side, which proved its excellent classification performance. In addition, SHAP is used to visually analyze the model to explore the influence of various factors on the severity of the accident. The results show that, for minor injury/serious injury/death, the most critical influencing factors are collision type, personnel category, and collision type. In order to avoid the occurrence of fatal accidents, we can pay more attention to and prevent unilateral collisions of passengers/trucks, and the accident is characterized by starting, running off course under the high speed limit road, colliding with the target vehicle, or hitting an obstacle. And should pay attention to the slack psychology of drivers who drive on the 30～60 km/h speed limit road.

Key words: Vehicle traffic accident; Traffic safety; Accident characteristics; Machine learning; Predictive model; Model interpretation

摘要： 为了研究影响机动车交通事故结果严重性的显著因素，掌握其发生特征与规律，将中 国事故深度调查（CIDAS）数据库中 2 966 条机动车事故进行建模，以包含人-车-路-环境的 19 个因素作为输入，将事故结果作为输出。在研究中引入包含准确率、查准率、召回率、F-1 分数等指标的评估体系，将 XGBoost 模型与 LightGBM、随机森林与 CatBoost 四种模型进行 并列对比，证明了其分类性能的优异。此外利用 SHAP 对于模型进行了可视化分析，探究各 种因素对于事故严重程度的影响。结果表明，对于轻伤/重伤/死亡而言，最关键的影响因素分 别为碰撞类型、人员类别、碰撞类型。为了避免死亡事故的发生，需重点关注注预防客/货车 单方碰撞事故、事故特征为起步/停止、高道路限速下跑偏/碰撞对象车辆/撞障碍物的场景， 并且应对于车辆行驶在 30～60 km/h 限速道路的驾驶员的懈怠心理重点关注。

关键词: 机动车交通事故；交通安全；事故特征；机器学习；预测模型；模型解释