School travel mode choice in two medium-sized south Asian cities: Cross-city transferability and explainable machine learning approaches

Trips for educational purposes represent a significant portion of morning and evening peak hour trips. These trips, if carried out by private transport, can lead to several negative consequences including increased traffic congestion, air and noise pollution, and driver discomfort. This study aimed at predicting the mode choices of school-going students in two medium-sized South Asian cities, Kandy, Sri Lanka, and Sahiwal, Pakistan. City-specific classification models were developed for each city, followed by cross-city evaluations using a subset of common features. SHapley Additive exPlanations (SHAP) were employed to interpret model behavior and assess the stability of learned decision logic across contexts. Ensemble models, particularly CatBoost and Gradient Boosting, consistently outperformed linear and single-tree classifiers in both cities, with substantially stronger predictive performance observed in Sahiwal due to richer household and contextual information. SHAP analyses reveal a shared behavioral foundation across cities in which cost-related variables dominate mode choice decisions. Higher costs for both private and sustainable modes are associated with continued reliance on the corresponding mode, indicating necessity-driven, mode-aligned behavior rather than cost-induced switching. Distance, income, and school type exert secondary but context-dependent effects within cities. Cross-city transferability analysis demonstrates limited and asymmetric generalizability. Models trained in one city experience pronounced performance degradation and systematic classification biases when applied to the other. SHAP-based diagnostics show that transferred models undergo marked reconfiguration of decision logic, including reduced spatial sensitivity and disproportionate reliance on cost signals, with evidence of decision-structure collapse under certain transfer directions. These results highlight the strong context dependence of school travel behavior and the need for locally calibrated, explainable modeling approaches.