Deep Neural Networks for Choice Analysis

Individual choice has been an enduring question across disciplines. Deep neural networks (DNNs) have demonstrated their high predictive power over the classical discrete choice models (DCMs) in many empirical studies. However, DNNs as a new modeling paradigm still present pressing challenges in interpretation, generalization, and robustness. This presentation introduces a deep choice framework that synergizes DNNs and DCMs to model individual travel decision. It demonstrates that the DNNs can provide economic information as complete as classical DCMs, including choice predictions, choice probabilities, market shares, substitution patterns of alternatives, social welfare, heterogeneous values of time, among many others, thus partially resolving the interpretation challenge. It introduces how to use the prior behavioral knowledge to design a particular DNN architecture with alternative-specific utility functions, which improves the generalizability of DNNs with a domain-knowledge-based regularization method. It then extends the framework to deep hybrid models, which integrates classical numerical data and the unstructured data (i.e., imagery and graphs) to analyze travel behavior. Overall, this presentation lays out a new foundation of using DNNs to analyze travel demand, enhancing economic interpretation, architectural design, and robustness of deep learning through classical utility theory.

SPEAKER

Shenhao Wang is an assistant professor and the director of the Urban Artificial Intelligence Laboratory at the University of Florida. He is also a research affiliate to Urban Mobility Lab and Media Lab at the Massachusetts Institute of Technology. He seeks to develop fundamental theory for urban science using artificial intelligence. He develops deep choice models, which analyze individual decision-making by integrating discrete choice models and deep learning with applications to urban travel behavioral analysis. He also analyzes collective mobility networks by integrating classical network theory and graph neural networks to quantify risk and uncertainty, thus promoting resilient economic growth. Dr. Wang completed his interdisciplinary Ph.D. in Computer and Urban Science at Massachusetts Institute of Technology in 2020. He received B.A. in Economics from Peking University (2014) and B.A. in architecture and law from Tsinghua University (2011), Master of Science in Transportation, and Master of City Planning from MIT (2017).