School travel mode choice in a secondary city of Pakistan: a comparison of multinomial logit and multilayer perceptron models

School travel mode choice remains comparatively underexplored, particularly in developing countries and in studies comparing conventional discrete choice and machine-learning approaches. To address this gap, this study compares the effectiveness and robustness of Multinomial Logistic Regression (MNL) and a Multilayer Perceptron (MLP) Neural Network in predicting and explaining students' mode choice in Abbottabad, Pakistan. A household-based questionnaire survey yielded 370 valid observations; 80% of the data were used for model training/calibration and 20% for validation. Slovin's formula was used to calculate the sample size. A multinomial logistic regression model was estimated using SPSS Statistics, while an MLP Neural Network was developed in Python for predictive evaluation. The logistic regression model observed that age, gender, travel length, travel time, travel cost, parking availability in the dwelling area, and the waiting time for public transport were important factors influencing mode choice behavior. In predictive evaluation, the MLP model outperformed the MNL model in classification accuracy and ROC–AUC, indicating better predictive reliability under class-imbalanced conditions. This study suggests incorporating the MLP model as a complementary tool alongside MNL for mode choice modeling in urban transportation systems and planning.