REFRAMING ALGORITHMIC FAIRNESS: A PARADIGM FOR FAIR, ACCURATE, AND FLEXIBLE MODEL DEVELOPMENT

Machine learning increasingly shapes high-stakes decisions, offering benefits like improved efficiency and accuracy—but also raising challenges when systems fail. Unlike human decision-makers, algorithms often lack clear paths for recourse or repair. This dissertation develops algorithmic tools for responsible and flexible machine learning, focusing on models that can adapt post-deployment, incorporate human feedback, and scale to real-world use. It introduces bias bounties--- a framework for users to collaboratively identify and patch model failures—and offers algorithms for post-processing models to meet fairness goals without retraining or rigid assumptions. It advances the theory of multicalibration, showing when fairness can be achieved without sacrificing accuracy, and proposes a simple swap-regret formulation of multicalibration which can be used in a variety of contexts, such as collaborative learning. Finally, it leverages multicalibration for flexible model development, demonstrating techniques to flexibly postprocess multicalibrated predictors to meet various downstream fairness objectives.

Author dc:creator

• Globus-Harris, Ira

Advisors dc:contributor.advisor

• Kearns, Michael
• Roth, Aaron

• Computer Sciences

Language dc:language.iso

en

Repository record dc:identifier.uri

https://repository.upenn.edu/handle/20.500.14332/61722

OAI identifier oai:identifier

oai:repository.upenn.edu:20.500.14332/61722