Fair Representation Learning with Controllable High Confidence Guarantees via Adversarial Inference
Yuhong Luo, Austin Hoag, Xintong Wang, Philip S. Thomas, Przemyslaw A. Grabowicz
TL;DR
This work tackles the problem of ensuring fair representations in downstream tasks by providing high-confidence guarantees for demographic parity. It introduces FRG, a framework that combines candidate selection, adversarial inference, and a fairness test to produce representations whose worst-case downstream DP disparity $\Delta_{\text{DP}}(\tau,\phi)$ remains below a user-specified threshold $\varepsilon$ with probability at least $1-\delta$. The method formalizes $g_{\varepsilon}(\phi) = \sup_{\tau} \Delta_{\text{DP}}(\tau,\phi) - \varepsilon$ and uses a 1-$(\delta)$ confidence upper bound $U_{\varepsilon}(\phi, D_f)$ derived from an optimal adversary to certify fairness, returning NSF when certification is not possible. Empirically, FRG demonstrates strong fairness guarantees across three real-world datasets while maintaining competitive predictive performance, outperforming six SOTA fair representation learning methods in consistently bounding DP disparity under adversarial evaluation. The work provides a practical, principled pathway for producing fair representations with explicit, tunable guarantees, highlighting limitations such as i.i.d. assumptions and reliance on approximate adversaries while outlining future extensions to other fairness notions and distributional shifts.
Abstract
Representation learning is increasingly applied to generate representations that generalize well across multiple downstream tasks. Ensuring fairness guarantees in representation learning is crucial to prevent unfairness toward specific demographic groups in downstream tasks. In this work, we formally introduce the task of learning representations that achieve high-confidence fairness. We aim to guarantee that demographic disparity in every downstream prediction remains bounded by a *user-defined* error threshold $ε$, with *controllable* high probability. To this end, we propose the ***F**air **R**epresentation learning with high-confidence **G**uarantees (FRG)* framework, which provides these high-confidence fairness guarantees by leveraging an optimized adversarial model. We empirically evaluate FRG on three real-world datasets, comparing its performance to six state-of-the-art fair representation learning methods. Our results demonstrate that FRG consistently bounds unfairness across a range of downstream models and tasks.