Seminar at TUe

On September 18th, 2024, I had the opportunity to present my on-going research work at Eindhoven University of Technology (TUe). Hosted by Assistant Professor Melvin Wong, the seminar focused on my work with generative latent class choice models, specifically using Variational Autoencoders (VAEs). Below are the title and abstract on my talk at TUe.

Title: Generative Latent Class Choice Models: The Case of Variational Autoencoders.

Abstract: The field of discrete choice modeling is undergoing significant advancements as researchers are increasingly using and comparing machine learning (ML) techniques to discrete choice model (DCM) performance. Moreover, ML techniques are being integrated with DCMs to allow for more accurate predictions and a deeper understanding of complex choice behaviors, all while maintaining the behavioral and economic interpretability that characterizes traditional DCMs. However, such efforts are still in the early stages, and while ML is a broad field that embody several learning paradigms, most of the hybrid ML-DCM models rely on supervised learning classifiers or unsupervised clustering and segmentation methods. Recently, deep generative models, a type of deep learning models that can generate new data points (e.g., text, video, images, or different types of synthetic data) that resemble the data they were trained on, have been creating waves and transforming many industries due to their impressive generative capabilities. However, deep generative models have yet to be embraced by the choice modelling community in hybrid frameworks that capitalize on their generative capabilities without giving up one the advantages of the random utility-based discrete choice models. This paper introduces a novel integration of Variational Autoencoders (VAEs), a type of generative model that combines neural networks with probabilistic graphical models, and Latent Class Choice Models (LCCMs). The proposed VAE-LCCM model leverages the generative and dimensionality reduction capabilities of VAEs to improve the estimation of class membership, enhance synthetic data generation, and better account for heterogeneity in choice behavior, all while preserving the behavioral and economic interpretability of traditional LCCMs. The model is applied to a case study, demonstrating improved goodness-of-fit and out-of-sample generalization compared to standard DCMs and ML approaches. The results suggest that the VAE-LCCM offers a promising direction for future research in discrete choice modeling by integrating the strengths of generative models and traditional econometric methods.