A Comprehensive Evaluation of Machine Learning Algorithms and Discrete Choice Model for Predicting Long Distance Travel Mode Preferences

Rapid urbanization in recent years in Bangladesh has significantly increased the prevalence of long-distance travel (LDT). Consequently, studying LDT mode choice modeling has become crucial in this context. However, most recent studies have primarily focused on either traditional Logit Models or Machine Learning models for classifying LDT mode choices, specifically between Intercity Train and Intercity Bus. This study aims to provide a comprehensive comparative analysis of Machine Learning models and the Discrete Choice Model for LDT mode choice classification. Drawing on extensive literature, 15 influential features were identified and utilized for classification. Data was collected from transportation hubs using both Revealed Preference (RP) and Stated Preference (SP) surveys. Eleven Machine Learning models, including Naïve Bayes (NB), K-Nearest Neighbor (KNN), Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), AdaBoost, CatBoost, Gradient Boost (GB), Passive Aggressive Classifier (PAC), Ridge Classifier, LightGBM, and one Discrete Choice Model, Binary Logit (BL), were compared. Results show that Random Forest emerged as the best-performing Machine Learning model, achieving accuracy scores of 0.86 and 0.87 before and after optimization, respectively. In contrast, the Binary Logit model achieved accuracy scores of 0.72 and 0.73. The Random Forest model also outperformed the Binary Logit model across other metrics, including Precision, Recall, F1 Score, and the Confusion Matrix. Furthermore, the AUC-ROC curve revealed a superior AUC value of 0.93 for the Random Forest model compared to 0.82 for the Binary Logit model. To identify dominant predictors, SHAP analysis was conducted for both Random Forest and Binary Logit models, further validating the Random Forest model's superiority in classification tasks.