Virginia Tech
Supervised Variational Autoencoders for Structural Learning and Statistical Inference with Heterogeneous Data
Abstract
dc:description.abstractLarge-scale datasets, such as images, often exhibit heterogeneous structures caused by diverse subpopulations or complex experimental designs. Extracting meaningful low-dimensional representations from such data, while accounting for heterogeneity and enabling statistical inference, remains a significant challenge. This dissertation addresses two related challenges within variational autoencoder (VAE) frameworks. The first project introduces the Generalized Variational Autoencoder (GVAE), a unified deep generative model that integrates dimension reduction, structural learning, and adaptive prediction into a single framework. GVAE constructs a composite latent space consisting of a Gaussian component for feature extraction and a stick-breaking process component for capturing latent subpopulation structure. A mixture-of-experts predictor linked to this composite latent space produces predictions that adapt to the subgroups. Simulation studies and an application to brain tumor MRI data demonstrate that GVAE achieves improved predictive accuracy over competing models while effectively recovering the underlying latent heterogeneous structure. The conditional generative component further reveals how images vary with the response, providing an additional tool for understanding complex data. The second project extends the VAE framework to association testing by proposing a two-step M-estimation approach in which VAE encoder parameters serve as first-step nuisance estimators for testing the association between high-dimensional inputs and a scalar response variable. We establish two Wilks-type asymptotic results depending on how the first-step nuisance parameters are obtained. When the nuisance parameters are trained with an unsupervised VAE, the classical Wilks' theorem holds and the two-step likelihood ratio statistic converges to a chi-squared distribution under the null. When the parameters are trained with a supervised VAE, the classical approximation may fail and we conjecture that a scaled likelihood ratio statistic still follows an approximate chi-squared distribution with adjusted degrees of freedom. Through simulation studies, we investigate the asymptotic behavior of the proposed test statistic and evaluate the type I error and power.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Statistics
- Department dc:contributor.department
- Statistics
- Grantor dc:publisher
- Virginia Tech
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Lee, Jaeyoung
- Chair dc:contributor.committeechair
-
- Zhu, Hongxiao
- Committee members dc:contributor.committeemember
-
- Kim, Inyoung
- Du, Pang
- Leman, Scott C.
Subjects
dc:subject × 5Rights
dc:rights- Statement dc:rights
-
- In Copyright
- Licence dc:rights.uri
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Other
- vt_gsexam:47415