Safe and Explainable Machine Learning for High-Impact Decisions

Abstract

My thesis develops methods for safe and explainable machine learning in high-impact domains, with a focus on fairness. I first addressed bias mitigation through multi-task learning and uncertainty estimation, balancing fairness and accuracy using Pareto optimization. I then developed a framework using large language models to improve causal discovery of bias pathways. My ongoing work focuses on translating high-level fairness policies into causal model constraints, enabling automated enforcement of legal fairness requirements in machine learning pipelines. This interdisciplinary approach bridges technical fairness methods with policy-aligned model design.