A1098
Title: MixTasteNet: A neural-embedded mixed logit model
Authors: Alvaro Gutierrez Vargas - KU Leuven (Belgium) [presenting]
Martina Vandebroek - K.U. Leuven (Belgium)
Michel Meulders - KU Leuven (Belgium)
Abstract: The MixTasteNet model is proposed, a novel hybrid of an Artificial Neural Network (ANN) and a Mixed Logit (MIXL) model, for modelling "taste heterogeneity" in Discrete Choice Models (DCM). While conventional Multinomial Logit (MNL) models can incorporate observed heterogeneity by including interactions between alternative-specific regressors and individual characteristics, it is a cumbersome trial-and-error process that can rapidly increase the number of estimated parameters. In contrast, random coefficient models, such as MIXL models, aim to capture unobserved heterogeneity through distributional assumptions over the taste parameters. Hybrid models use individuals' characteristics to feed an ANN and produce heterogeneous taste parameters that are included in the utility specification. The MixTasteNet model goes a step further by simultaneously including random coefficients, which capture unobserved heterogeneity, and an ANN component, which captures the observed heterogeneity, in the utility specification. Notably, the proposed model is the first hybrid specification used in DCM that incorporates random coefficients and an ANN to model individuals' preferences. Finally, the MixTasteNet model accurately recovers the true models' parameters while achieving the predictability of the ground-truth model was demonstrated via simulation.
