Neural-embedded discrete choice models

Abstract

This dissertation is motivated by the possible value of integrating theory-based discrete choice models (DCM) and data-driven neural networks. How to benefit from the strengths of both is the overarching question. I propose hybrid structures and strategies to flexibly represent taste heterogeneity, reduce potential biases, and improve predictability while keeping model interpretability. Also, I utilize neural networks' training machinery to speed up and scale up the estimation of Latent Class Choice Models (LCCMs). First, I embed neural networks in DCMs to enable flexible representations of taste heterogeneity and enhance prediction accuracy. I propose two neural-embedded choice models: TasteNet-MNL and nonlinear-LCCM. Both models provide a flexible specification of taste as a function of individual characteristics. TasteNet-MNL extends the Multinomial Logit Model (MNL).

A feed-forward neural network (TasteNet) is utilized to predict taste parameters as a nonlinear function of individual characteristics. Taste parameters generated by TasteNet are further fed into a parametric logit model to formulate choice probabilities. I demonstrate the effectiveness of this integrated model in capturing nonlinearity in tastes without a priori knowledge. Using synthetic data, TasteNet-MNL is able to recover the underlying utility specification and predict more accurately than some misspecified MNLs and continuous mixed logit models. TasteNet-MNL also provides interpretations close to the ground truth. In an application to a public dataset (Swissmetro), TasteNet-MNL achieves the best out-of-sample prediction accuracy and discovers a broader spectrum of taste variation than the benchmark MNLs with linear utility specifications. Nonlinear-LCCM enriches the class membership model of a typical LCCM.

I represent an LCCM by a neural network and add hidden layers with nonlinear transformations to its class membership model. The nonlinearity introduced by the neural network provides a flexible approximation of the mixing distribution for both systematic and random taste heterogeneity. I apply this method to model Swissmetro mode choice. The nonlinear-LCCM outperforms an LCCM with a linear class membership model with respect to the out-of-sample prediction accuracy. Nonlinear-LCCM also provides interpretable taste parameters for each latent class.