A post-processing framework for group fairness

Machine learning models are increasingly powering automated decision-making systems that influence everyday life, thanks to their ease of deployment. But this convenience belies the risk that, without proper oversight, they may cause disparate impacts across demographic groups. For instance, models trained on data shaped by historical inequalities can propagate those biases and disadvantage protected groups: ProPublica’s analysis of the COMPAS recidivism tool showed that it disproportionately mislabeled Black defendants as high risk. Group fairness definitions, including statistical parity and equal opportunity, formalize such disparities and form the basis of many fair learning algorithms. However, the effectiveness of these algorithms is often hindered by practical challenges. When direct access to sensitive attributes is legally restricted, fair algorithms typically rely on proxy predictors to infer these attributes; if the proxies are miscalibrated, fairness guarantees may be invalidated. This is an instance of the more general issue of distribution shift between training and test environments. In addition, privacy constraints, notably differential privacy, require injecting random noise to protect individual data, but it obscures the group statistics needed to enforce fairness. These challenges all fall under the broader goal towards trustworthy machine learning—building models that are reliable, safe, and ethically sound—but they are often addressed in isolation, resulting in methods that are not necessarily compatible. We propose a unified algorithmic framework for learning fair classifiers that also addresses robustness and privacy cohesively. At its core is a post-processing algorithm that applies lightweight adjustments to the predictions output by existing models to achieve group fairness; it recovers the optimal fair classifier when the base predictor is Bayes-optimal. The model-agnostic and post-hoc nature makes it particularly well-suited to emerging paradigms that derive predictors from pre-trained models rather than training from scratch, such as via prompting large language models. To address distribution shift, we introduce algorithms that calibrate the base predictor to the test distribution before post-processing, or, when the shift is unknown, a robust procedure against worst-case shifts within an uncertainty set. To satisfy differential privacy, we use the fact that post-processing depends on the training data only through the empirical joint distribution of model outputs and sensitive attributes, and ensure privacy simply by substituting in a private estimate. We provide open-source code to support the practical adoption of fairness mitigations, and analyses that explore the tradeoffs between accuracy, fairness, robustness, and privacy.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Xian, Ruicheng

Contributors dc:contributor

• Zhao, Han
• Banerjee, Arindam
• Tong, Hanghang
• Roth, Aaron
• Kamath, Gautam

• group fairness
• algorithmic fairness
• machine learning

Statement dc:rights

• Copyright 2025 Ruicheng Xian

Language dc:language

en

Handle dc:identifier

https://hdl.handle.net/2142/132546

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/132546