COMMITTEE CHAIR: Dr. Mohamed Chouikha

TITLE: ACHIEVING EQUITY IN COMPUTATIONAL LEARNING SYSTEMS THROUGH FAIR OPTIMIZATION AND PRIVACY-PRESERVATION

ABSTRACT: Bias in machine learning systems continues to raise critical concerns about equity, accountability, and reliability-especially in socially consequential domains such as criminal justice, housing finance, and healthcare. While fairness-aware learning has introduced important metrics and mitigation techniques, most existing approaches remain siloed, addressing bias at isolated points in the machine learning pipeline without accounting for systemic interdependencies between fairness, privacy, and model performance. This dissertation addresses that gap by advancing a unified, technically grounded framework for bias mitigation that spans the pre-processing, in-processing, and post-processing phases of the Machine Learning lifecycle. Building upon the foundations of fairness constraints and bias detection, this work introduces new contributions that strengthen both theoretical understanding and practical implementation. These include a class of interpretable, non-linear models, stochastic generative architectures such as tabular diffusion models for fairness-aware synthetic data generation, and a novel harm-aware metric that quantifies the societal impact of false positive errors across groups. In addition, the study proposes a mathematical audit mechanism that formalizes deviation from ideal fairness baselines. Experimental validation is conducted on four domain-relevant datasets to evaluate the effectiveness of the framework. Results show improved group fairness metrics with minimal degradation in accuracy and F1 scores, along with meaningful reductions in membership inference risk, highlighting the promise of synthetic data as a privacy-preserving intervention. These findings emphasize that fairness in Machine Learning cannot be achieved solely through isolated algorithmic tweaks but must be pursued through cohesive, context-aware, and ethically informed system design. This dissertation contributes a foundation for building fair and resilient machine learning pipelines and provides actionable insights for developing trustworthy Artificial Intelligence systems that are both generalizable and socially responsible.

Room Location: Electrical and Computer Engineering Department Conference, Room 315D