Comparing Choice Model and Machine Learning for Commuter Behavior Prediction: Evaluating Model Transferability Across Urban and Cultural Contexts

Understanding commuter behavior is essential for developing effective and sustainable urban transportation systems. As travel patterns evolve due to infrastructure changes, emerging mobility services, and shifting commuter preferences, accurately predicting travel choices becomes increasingly challenging. This study compares five predictive models—Multinomial Logit (MNL), Random Forest (RF), XGBoost, CatBoost, and a Neural Network (NN)—to evaluate their accuracy and transferability across changing urban mobility conditions. We use commuter survey data collected in 2023 from employees living and working in the city of Rome to train and validate the models via 5-fold cross-validation (CV), followed by ANOVA tests to assess statistical differences in performance. To examine transferability, we apply the trained models to two out-of-sample datasets: one from Rome in 2022, capturing post-COVID travel behavior changes within the same urban context, and one from Turin in 2023, representing a distinct urban and cultural setting. Results show that tree-based ensemble models such as XGBoost and CatBoost achieve the highest predictive accuracy and also demonstrate strong generalization across both temporal and spatial contexts. In contrast, the MNL model offers greater interpretability but lower performance, and the neural network underperforms in transferability. These findings highlight the trade-offs between predictive power, robustness, and model transparency in transportation modeling, offering valuable guidance for data-driven transportation planning.