Decomposing random taste heterogeneity in discrete choice modeling: A mixture density network-embedded logit model

While advanced discrete choice models such as the mixed logit (MXL) and neural-embedded logit models capture taste heterogeneity, important limitations remain. MXL models rely on pre-specified parameter distributions, whereas neural-embedded models typically produce point estimates of taste parameters conditioned on demographics, which fail to flexibly represent random heterogeneity within demographic groups. To bridge this research gap, this study develops a robust Mixture Density Network-embedded Logit (MDN-Logit) model that extends the point estimate of taste parameters to full distributional representations. The model uses a mixture density network to learn flexible, data-driven mixture distributions of taste parameters conditioned on individual socio-demographic characteristics, thereby jointly capturing systematic and random heterogeneity within a unified framework. The MDN-Logit model is validated using synthetic data, the Swissmetro stated preference dataset and a revealed preference freight dataset. Results demonstrate that it successfully recovers complex parameter distributions along with accurate value of time and distance sensitivity estimates, while achieving superior goodness-of-fit and predictive performance over benchmark models such as MXL and TasteNet. Ultimately, its ability to compute policy-relevant indicators, including the value of time and elasticities, makes it a powerful tool for granular passenger and freight transportation planning.