PhD Defense by Sung Hoo Kim
Ph.D. Thesis Defense Announcement
An Integrated Framework for Exploring Finite Mixture Heterogeneity in Travel Demand & Behavior
Sung Hoo Kim
Dr. Patricia L. Mokhtarian (CEE)
Dr. Giovanni Circella (CEE), Dr. Jorge A. Laval (CEE),
Dr. Shatakshee Dhongde (ECON), Dr. Joan L. Walker (CEE, UC Berkeley)
Date & Time: June 28, 2021 | 1:00 PM EST
In recent years we have faced a plethora of social trends and new technologies such as shared mobility, micro-mobility, and information and communication technologies, and we will be facing many more in the future (e.g. self-driving cars, disruptive events). In this context, the perennial mission of transportation behavior analysts and modelers - to model behavior/demand so as to understand behavior, help craft responsive policies, and accurately forecast future demand - has become far more challenging.
Specifically, behavioral realism and predictive ability are two key goals of modeling (travel) behavior/demand, and a key strategy for achieving those goals has been to introduce some type of heterogeneity in modeling. Thus, this thesis aims to improve our behavioral modeling by accounting for heterogeneity, with clues from the ideas of data/market segmentation, finite mixture, and mixture modeling. The objectives of the thesis are: (1) to build a framework for modeling finite mixture heterogeneity that connects seemingly less related models and various methodological ideas across domains, (2) to tackle various heterogeneity-related research questions in travel behavior and thus show the empirical usefulness of the models under the framework; and (3) to examine the potential, challenges, and implications of the framework with conceptual considerations and practical applications.
Five inter-related studies in this thesis illuminate some part(s) of the framework and delineate how key concepts in the framework are connected to each other. (a) The thesis overviews the topics of heterogeneity and mixture modeling in transportation and provides the landscape and details of how we have used mixture modeling. (b) Extending the idea of a finite segmentation approach, the thesis connects and compares three models for treating finite-valued parameter heterogeneity: deterministic segmentation, endogenous switching, and latent class models. The study discusses their similarities and differences from conceptual and empirical standpoints. (c) The thesis explains the confirmatory latent class approach and its potential usefulness, as opposed to the conventional exploratory approach. Adopting this perspective, the study embraces zero-inflated models under the confirmatory latent class approach and demonstrates their empirical value. (d) The thesis introduces the idea of combining latent class and endogenous switching models. Conceptual and empirical differences between the standard latent class model and the proposed approach are discussed. (e) The dissertation illuminates the linkage between finite mixture modeling (specifically in “indirect application”) and the mixture of experts (MoE) architecture, introduced in machine learning. The study proposes to use MoE as a data-driven exploratory tool to capture nonlinear/interaction effects (which are types of parameter heterogeneity), and exhibits its ability using synthetic and empirical data. The thesis concludes with discussions about challenges, potential technical advances, and outlook for the framework.
The dissertation is expected to give conceptual/methodological insights on the framework for modeling finite mixture heterogeneity and how various methodologies are connected under the framework. As well, the studies provide rich discussions about study-specific empirical findings and their implications. Thus, the dissertation can help improve our behavior/demand models by serving as a navigational compass for analysts.
- Workflow Status: Published
- Created By: Tatianna Richardson
- Created: 06/11/2021
- Modified By: Tatianna Richardson
- Modified: 06/11/2021