SHAP-based Ensemble Learning for Assessing Key Factors Influencing Intention to Use Intercity Railway Service

This study identifies and ranks key factors influencing the intention to use intercity railway services in Thailand by employing advanced machine learning algorithms and interpretability techniques. Six algorithms—Random Forest, Gradient Boosting, AdaBoost, XGBoost, LightGBM, and CatBoost—were used to build predictive models, with data split 85:15 for training and testing, and hyperparameter tuning for robustness. XGBoost achieved the highest accuracy (84.17%) and AUC (96%), closely followed by CatBoost. Class-wise, XGBoost had the highest F1-scores for “Neutral intention” (90.16%) and “High intention” (64.86%), while CatBoost demonstrated the highest precision for “No intention” (55.56%). SHAP (SHapley Additive exPlanations) was utilized for model interpretability, revealing that ease of trip planning and polite staff behaviour are the top determinants of high usage intention. XGBoost's superior predictive performance, coupled with SHAP's ability to provide transparent and interpretable insights, offers a powerful tool for policymakers and railway operators.