Comparing gradient boosting and discrete choice models for urban work-trip mode choice: a case study of Tiruchirappalli, India

Understanding urban work-trip mode choice is central to effective transport planning, yet accurately capturing complex travel behaviour remains challenging for conventional modelling approaches. This study presents a systematic comparison between traditional discrete choice models and gradient boosting–based machine learning models for predicting work-trip mode choice, using 11,270 revealed-preference trip observations from Tiruchirappalli, a medium-sized Indian city. Multinomial Logit, Probit, and Multinomial Logistic Regression models are benchmarked against three gradient boosting algorithms XGBoost, LightGBM, and CatBoost within a harmonised modelling framework employing a common set of explanatory variables and consistent data partitions. Model performance is evaluated using likelihood-based goodness-of-fit measures for discrete choice models and out-of-sample classification metrics, including accuracy, macro-averaged precision, recall, F1-score, and one-vs-rest ROC-AUC, for machine learning models. Results indicate that gradient boosting models, particularly LightGBM, achieve higher predictive accuracy and improved class-wise performance under severe modal imbalance, while discrete choice models provide compact behavioural representations with interpretable parameter estimates and stable goodness-of-fit measures. To address interpretability limitations of machine learning models, SHapley Additive exPlanations (SHAP) and partial dependence plots are employed, revealing that key determinants such as trip distance, travel cost, income, and vehicle ownership exhibit directional consistency with econometric coefficient estimates, while also exposing non-linearities and interaction effects not captured by linear utility specifications. The findings demonstrate that discrete choice and gradient boosting models offer complementary strengths: econometric models support behavioural interpretation and policy sensitivity analysis, whereas explainable machine learning models enhance predictive performance and diagnostic insight under complex and imbalanced data conditions. The study provides empirical evidence to inform the appropriate use of both modelling paradigms in urban work-trip travel demand analysis, particularly in medium-sized cities of the Global South.