Railway Accident Classification Prediction and Causal Analysis Based on Explainable Machine Learning

Applying data mining techniques to historical railway accident data for accident classification prediction and causal analysis holds significant practical importance for enhancing railway safety. A railway accident risk feature system is developed based on international railway accident reports and used as input variables for the Extreme Gradient Boosting (XGBoost) prediction model. The Harris Hawk Optimisation (HHO) algorithm is improved by introducing Latin Hypercubic Sampling, Adaptive Energy Factor and Elite Contrastive Learning Mechanism to propose the IHHO algorithm with stronger global exploration and local exploitation capabilities. Then, the IHHO algorithm is used to optimise the key hyperparameters of the XGBoost classification model to construct a railway accident classification prediction model. Finally, the SHAP method is applied for model interpretability analysis, identifying the key factors influencing the occurrence of different types of railway accidents. Experimental results show that the IHHO-optimized XGBoost model exhibits excellent performance in accident classification tasks, with high accuracy and stability.