基于组合模型的交通事故严重程度预测方法

由于各个单一分类模型对道路交通事故严重程度预测的局限性，本文致力于建立一种组合模型。本文结合卷积神经网络提取时空维度中的特征信息，采用stacking方式将CNNs与XGBoost组合，最终生成道路交通事故严重性的分类模型（多层提升方法），实验结果表明，此模型在测试集上预测精度为91.51%，组合模型比单一分类模型具有更好的分类结果。基于组合模型的分类结果，对交通事故特征进行重要性排序，开展特征相关性分析，为减少道路交通事故及减轻道路交通事故严重等级的管理措施提供参考依据。

A Traffic Accident Prediction Approach Based on An Ensemble Approach

As the limitations of the single classifier on traffic accident severity, an ensemble approach is proposed. Using CNNs to extract the features from the spatial dimension, getting an ensemble approach with XGBoost and CNNs by stacking (Multi-level Boosting Algorithm). The predicting precision of the approach is 91.51% on the validation set. In comparison with the single classification model, the result of the experiment shows a better performance. For providing useful information for reducing the number of traffic accidents and downgrading the severity of traffic accident, the paper gives out a correlation analysis by sorting the features based on the predictions.