基于互信息贝叶斯网络的交通事故严重程度分析

为掌握省际客运行业事故严重程度影响因素, 采用互信息及贝叶斯网络方法构建模型, 分析各因素变化与事故严重程度的定量互动关系。鉴于行业样本量较小及专家知识建模存在主观性, 采用改进离散算法挖掘数据, 提出结合互信息与交叉验证的先验网络构造方法。以上海市2005—2019年741起省际客运事故数据为例进行模型分析。结果表明: 对事故最敏感的影响因素为驾驶员性别、天气和车辆类型; 其中“女性驾驶员”“雪、大风、雾”“中型客车”对事故严重性的权重占比分别为13.5%, 8.8%和5.7%;此外, 驾驶员年龄对群死群伤事故贡献较小; 客车尺寸与安全性非单调关系; 00:00—05:00引发7人以上受伤的概率同比上升9%;季节、天气、时间因素与财产损失无直接关联。模型泛化能力优于对比模型, AUC均值为0.644 588, 命中率达到97.3%。 

An Analysis of Traffic Accident Severity Based on Mutual-information Bayesian Network

The methods of mutual information and Bayesian network are conducted to develop a model to grasp the factors affecting the severity of accidents in the inter-provincial bus industry. The quantitative interaction between changes in factors and the severity of accidents are analyzed. Given the limitation of the samples' size of the industry and the subjectivity of experts' knowledge of modeling, an improved discrete algorithm is used for data mining.A primary network construction method combining mutual information and cross-validation is proposed. Taking model analysis with 741 inter-provincial bus accidents in Shanghai from 2005 to 2019 as a case study, the results show that the most sensitive influencing factors of accidents are gender, weather, and vehicle type."Female driver""snow, wind, and fog""medium-size bus"account for 13.5%, 8.8%, and 5.7% of the weight of the accidents, respectively. Additionally, drivers' age has little contribution to the misfortune of group death and injury. Bus size has non-monotonic relationships with safety. The probability of more than seven people being injured during 00:00 to05:00 rises by 9%. The factors of season, weather, and time are not directly related to property loss. The generalization ability of the constructed model is better than other comparable models. The average AUC is 0.644 588, and the hit rate reaches 97.3%.