Research on predicting the severity of traffic accidents with unbalanced crash data in urban scenarios

To explore the impact of unbalanced traffic accident data on the prediction accuracy of the model in urban scenarios, rear-end crashes involving two vehicles in Chicago from 2016 to 2020 were selected for modeling, a “generative balance-machine learning prediction” framework for urban traffic safety analysis is constructed. The crash data was randomly split into training and testing sets in a 7:3 ratio. The training data was balanced using 1the Conditional Tabular Generative Adversarial Networks (CTGAN), which generates high-quality artificial crash data. The CTGAN-generated data was merged with the original training data to achieve data balance. An XGBoost model was then trained using the balanced training data and evaluated using the imbalanced testing data to assess its predictive performance. To compare the effectiveness of different data balancing algorithms, four commonly used oversampling and combination sampling algorithms were also employed. The results demonstrate that the artificial crash data generated by CTGAN exhibited better quality than other balancing algorithms. Additionally, the crash severity classification model trained on the CTGAN-balanced data outperformed other algorithms in terms of metrics such as G-mean, Balanced Accuracy, and the prediction accuracy for minor class crashes. Subsequently, SHAP was used to interpret the model's predictive results. It was found that the Causal action (lead vehicle), lighting conditions, Gender (lead vehicle), Cross-section type and Pavement condition had more prominent effects on serious injury and death accidents. The results show that CTGAN combined with XGBoost can obtain robust performance considering the whole and minority classes in urban unbalanced traffic safety data, providing data support and decision-making basis for urban fine safety governance.