Road Traffic Accident Severity Prediction under Unbalanced Data

In recent years, road traffic safety has been a major concern. Research on road traffic accidents is crucial to address people’s worries about their daily commutes. This study focuses on California, using U.S. traffic accident data from February 2016 to the end of 2021. The research employs two data equalization methods: single sample and mixed sample. Traffic accident severity prediction models are established based on the XGBoost algorithm. These models are compared using accuracy and ROC curves. The model with the "random undersampling + SMOTENC oversampling" sequence is the most effective. The SHAP interpretability method analyzes accident factors’ impact on the model results. Consequently, specific strategies for preventing road traffic accidents are proposed.