Contributions to Machine Learning with Abstention and Surrogate Modeling with Complex Outputs

TR Number

Date

2026-06-16

Journal Title

Journal ISSN

Volume Title

Publisher

Virginia Tech

Abstract

Machine learning and statistical modeling have become integral components of modern scientific and engineering analysis, providing powerful tools to model and understand complex phenomena. However, real-world applications frequently present unique challenges, such as data imbalance, fairness constraints, and high-dimensional structured outputs. This dissertation presents novel statistical methodologies to address these challenges, with a focus on classification problems under data imbalance and fairness constraints, efficient emulation of computer experiments with tensor-valued outputs, and simulation-based experimental design and analysis. Chapter 1 presents the motivation and research objectives. Chapter 2 develops methods for logistic classification with a rejection option, addressing challenges of data imbalance and fairness constraints with respect to sensitive attributes. The proposed approach generalizes existing methods by incorporating convex surrogate loss functions and fairness-aware optimization. Performance is evaluated through comprehensive simulation studies and real-world datasets, including a case study on Methicillin-resistant Staphylococcus aureus (MRSA) prediction using imbalanced electronic health record (EHR) data. Chapter 3 introduces a so-called Deco-GPT method, a Gaussian process-based emulator for computer experiments with tensor-valued outputs. The method applies tensor unfolding and higher-order singular value decomposition (HOSVD) to improve scalability and predictive accuracy. An extension for handling missing output values is incorporated via matrix completion. Performance is evaluated through simulations and case studies, demonstrating strong performance across diverse settings. Chapter 4 introduces a simulation framework using CARLA to generate structured datasets, including safety-critical driving events and a route completion dataset with mixed inputs. We consider deep Gaussian process models for analyzing such data with mixed inputs. More specifically, we handle qualitative factors via indicator-based representation and learned embeddings, and compare against standard Gaussian process and latent variable Gaussian process models. Chapter 5 summarizes the main contributions of this research and discusses potential directions for future work.

Description

Keywords

Classification with Rejection, Imbalanced Data, Fairness Constraints, Gaussian Process, Tensor-Valued Outputs

Citation