考虑数据不平衡的城市道路乘用车致命事故率分析

An Analysis of Fatal Accident Rates of Passenger Cars on Urban Roads Considering Imbalanced Data Samples

城市道路交通事故频发,而事故数据存在明显不平衡,不同因素间的耦合作用对城市道路乘用车致命事故率分析造成极大挑战。为此提出了1种集成重采样、贝叶斯网络(Bayesian networks,BN)和关联规则(association rule method,ARM)的三阶段事故率分析方法。基于国家事故深度调查体系的1105例城市道路乘用车事故数据,从驾驶人、车辆、道路、环境这4个方面选取16个潜在特征变量构建BN模型;鉴于数据不平衡时会导致BN模型性能下降的问题,提出在构建BN模型前利用合成少数类过采样技术(Synthetic Minority Over-sampling Technique,SMOTE)和聚类中心进行数据重采样,并比较分析各类采样技术下不同BN模型的综合性能;基于最优BN模型并结合ARM,推理不同影响因素及因素的耦合作用对致命事故率的影响。结果表明:重采样方法可以显著提升BN模型的综合性能,以及识别风险因素的能力。其中SMOTE采样技术结合GTT算法构建的BN模型的AUC最高,达0.793。此外,相较于原始不平衡数据构建的BN模型,经SMOTE采样后构建的BN模型多挖掘了6个风险因素;“机动二/三轮车”与“超速行驶”耦合时致命事故率最高,达80.4%。“机动二/三轮车”与“存在视野盲区”耦合时,致命事故率达77.4%;乘用车在四枝分叉口左转时,容易与汽车发生碰撞,但致命事故率低于20%。本方法能够降低数据不平衡对道路交通事故分析的影响,并实现风险因素的耦合作用分析,进而预防和降低城市道路致命事故的发生。

Traffic accidents on urban roads are frequent,and there is a significant imbalance in accident data.The coupling between different factors caused great challenges in analyzing the fatal accident rate of passenger vehicles on urban roads.Therefore,a three-stage method that integrating resampling,Bayesian networks(BN)and association rule method(ARM)is proposed.Based on the data of 1105 urban road passenger car accidents from the National Automobile Accident In-Depth Investigation System(NAIS),the BN model is constructed by selecting 16 poten tial feature variables from four aspects:driver,vehicle,roadway and environment.Considering the problem that the imbalance of accident types can lead to the degradation performance of BN model.Proposed data re-sampling using Synthetic Minority Over-sampling Technique(SMOTE)and Cluster Centroids(CC)before the construction of BN model.Compare the comprehensive performance of different BN models under various sampling techniques.Finally,based on the optimal BN model and combined with the ARM,the effects of different influencing factors and the coupling effect of factors on the fatal accident rate were analyzed.The results show that re-sampling method can significantly improve the comprehensive performance of BN models and the ability to identify risk factors.Among them,the BN model constructed by SMOTE sampling technique combined with GTT algorithm has the highest AUC of 0.793.Besides,compared with the BN model constructed by the original imbalanced data,the BN model constructed by SMOTE sampling explores six more risk factors.The highest fatal accident rate was 80.4%when“motorized two/three-wheelers”are coupled with“speeding”.The next highest fatal accident rate is 77.4%when“motorized two/three wheelers”is coupled with“blind spots in the field of vision”.Passenger cars are prone to crash with cars when turning left at the Four-Way Intersection,but the fatal accident rate is less than 20%.This method can reduce the influence of data imbalance on the analysis of road traffic accidents,and realize the analysis of the coupling effect of risk factors,thus preventing and reducing the occurrence of fatal accidents on urban roads.