Predicting and Analyzing Accident Severity with Improved LightGBM Method

This paper proposes an improved light gradient boosting machine model for traffic accident severity prediction and analysis, which addresses the problems of class imbalance and poor interpretability. The model uses the CU_SMOTE sampling algorithm and the focal loss function using the whale optimization algorithm. The model is validated by real accident data and the factors of fatal accidents are analyzed by Shapley Additive Explanations. The results show that the model outperforms random forest and light gradient boosting machine in accuracy and recall, and that the absence of safety devices, accidents on regular roads, and frontal collisions are the main causes of fatal accidents.