Fairness-aware unsupervised learning methods for healthcare applications Zhuoping Zhou

Format:

Book

Thesis/Dissertation

Author/Creator:

Zhou, Zhuoping, author.

Contributor:

University of Pennsylvania. Applied Mathematics and Computational Science., degree granting institution.

Language:

English

Subjects (All):

Computer science.

Bioinformatics.

Oncology.

Demography.

0984.

0715.

0992.

0800.

0938.

Local Subjects:

Computer science.

Bioinformatics.

Oncology.

Demography.

0984.

0715.

0992.

0800.

0938.

Genre:

Academic theses

Physical Description:

1 online resource (141 pages)

Contained In:

Dissertations Abstracts International 87-12A

Place of Publication:

Ann Arbor : ProQuest Dissertations and Theses, 2026

Language Note:

English

Summary:

Fairness in machine learning has become critical for healthcare applications, where biased models risk perpetuating health disparities across demographic groups. Although extensive research addresses fairness in supervised learning, the challenge of achieving equitable outcomes in unsupervised learning-tasks such as dimensionality reduction, structure learning, and graphical model estimation-remains largely unexplored. This dissertation develops a principled framework for fairness-aware unsupervised learning and introduces four methodological contributions spanning canonical correlation analysis and graphical models.First, we propose Multi-Group Tensor Canonical Correlation Analysis (MG-TCCA), which discovers group-specific latent representations from tensor-valued data while imposing cross-group consistency through dual sparsity regularization. Second, we develop Fair Canonical Correlation Analysis (Fair CCA), formulating fairness as a multi-objective optimization problem on Riemannian manifolds that balances correlation maximization with equitable group performance. Third, we extend this framework to the multimodal setting through Fair Multimodal CCA (Fair MCCA), enabling fairness-aware joint analysis of more than two data modalities. Fourth, we introduce fairness-aware graphical model estimation, proposing Fair GLasso, Fair Covariance Graph, and Fair Binary Network methods that ensure equitable estimation of conditional independence structures across demographic groups.Central to our approach are two novel fairness metrics: the Correlation Disparity Error for CCA-based methods and the Graph Disparity Error for graphical models, both quantifying performance gaps across demographic groups. All proposed methods employ multi-objective optimization to achieve Pareto-optimal trade-offs between model accuracy and group fairness, with provable convergence guarantees. Experiments on the Alzheimer's Disease Neuroimaging Initiative (ADNI) and The Cancer Genome Atlas (TCGA) demonstrate 46-96% reductions in fairness disparity while preserving competitive statistical performance

Notes:

Source: Dissertations Abstracts International, Volume: 87-12, Section: A.

Advisors: Shen, Li; Long, Qi Committee members: Jin, Jin

Ph.D. University of Pennsylvania 2026

Vendor supplied data

Local Notes:

School code: 0175

ISBN:

9798247979159

Access Restriction:

Restricted for use by site license