Date of Award
2026-05-01
Degree Name
Doctor of Philosophy
Department
Data Science
Advisor(s)
Xiaogang Su
Abstract
High-dimensional logistic regression is particularly vulnerable to outliers, which can distort estimation, weaken variable selection, and harm prediction accuracy. This dissertation introduces RobENetLog, a robust elastic net logistic regression method that jointly selects informative variables and reliable observations by extending ideas from least trimmed squares regression to the classification setting. The proposed approach combines adaptive outlier detection with trimmed penalized estimation so that it can address contamination in both predictors and responses across a wide range of dimensional settings. Simulation studies show that RobENetLog consistently improves predictive performance and robustness in the presence of outliers, while remaining competitive with standard methods on clean data. The method also adapts automatically to different data structures, placing emphasis on the most informative type of outlier detection for the setting at hand. In addition, applications to three real-world datasets from chemometrics, genomics, and biomedical voice analysis demonstrate its practical value in challenging classification problems involving high dimensionality, correlated predictors, and suspected outliers. Across these studies, the method provides accurate classification, stable feature selection, and meaningful identification of atypical observations. An accompanying R package will make the approach readily available for practical applications.
Language
en
Provenance
Received from ProQuest
Copyright Date
2026-05
File Size
207 p.
File Format
application/pdf
Rights Holder
Chitra Bahadur Karki
Recommended Citation
Karki, Chitra Bahadur, "Robust Logistic Regression With Elastic Net For Contaminated High-Dimensional Data" (2026). Open Access Theses & Dissertations. 4706.
https://scholarworks.utep.edu/open_etd/4706