Improving Travel Mode Choice Predictions for Minority Modes with a Conditional Variational Autoencoder to Address Class Imbalance

Class imbalance in travel mode choice data represents a significant challenge to transportation modeling, often leading to biased parameter estimates and high prediction errors for less frequently chosen alternatives. This study presents a comparative analysis of the multinomial logit (MNL) model’s predictive performance when class imbalance is addressed using random oversampling (ROS), synthetic minority oversampling technique (SMOTE), class-weighting, and the conditional variational autoencoder (CVAE). To the best of our knowledge, this is the first application of CVAE to address class imbalance in travel mode choice prediction. Using three well-known datasets, we estimate the MNL model on both imbalanced and balanced data. Our results show that CVAE improves predictive performance only under severe class imbalance and only in one of three datasets, while performing worse than the unbalanced baseline in another. More critically, CVAE introduces sign reversals in key behavioral parameters in two of the three datasets. These reversals are not inherited from the original data but are specific to the CVAE augmentation method, and they render the resulting model economically implausible for policy analysis. We trace these reversals to a concrete mechanism: the distortion of attribute gaps between chosen and unchosen alternatives in the synthetic data. In contrast, ROS and class-weighting preserve behavioral plausibility while providing competitive predictive performance. These findings highlight a fundamental trade-off between predictive accuracy and behavioral interpretability under generative data augmentation. CVAE should be restricted to purely predictive applications, and its use must be accompanied by verification of coefficient coherence before any policy interpretation is attempted.