Experimentation of learning multinomial logit for travel mode choice: systematic evaluation of prediction, interpretability, and behavioural reliability

Neural-embedded discrete choice models (DCMs) combine behavioural structure with neural-network flexibility, but how neural-component flexibility affects predictive and behavioural performance remains unclear. This study uses the Learning Multinomial Logit (L-MNL) framework to develop three variants with progressively greater neural flexibility and evaluates them alongside multinomial logit (MNL) and deep neural network (DNN) models. A unified framework assesses predictive performance, interpretability, and behavioural reliability across 1,200 controlled experiments on three real-world datasets. Results show that greater neural flexibility can improve best-case predictive performance but also increases sensitivity to experimental settings and performance variability. MNL retains advantages in small-sample settings, whereas DNN performs better on larger datasets but shows weaker behavioural reliability. L-MNL-based models provide an intermediate balance between predictive flexibility and behavioural consistency. These findings support model selection under different data conditions and highlight the need to evaluate both predictive and behavioural dimensions in travel behaviour modelling.