Achievable Fairness on Your Data With Utility Guarantees

TL;DR

This work presents a computationally efficient approach to approximate the fairness-accuracy trade-off curve tailored to individual datasets, backed by rigorous statistical guarantees and introduces a novel methodology for quantifying uncertainty in the authors' estimates, thereby providing practitioners with a robust framework for auditing model fairness.

Abstract

In machine learning fairness, training models that minimize disparity across different sensitive groups often leads to diminished accuracy, a phenomenon known as the fairness-accuracy trade-off. The severity of this trade-off inherently depends on dataset characteristics such as dataset imbalances or biases and therefore, using a uniform fairness requirement across diverse datasets remains questionable. To address this, we present a computationally efficient approach to approximate the fairness-accuracy trade-off curve tailored to individual datasets, backed by rigorous statistical guarantees. By utilizing the You-Only-Train-Once (YOTO) framework, our approach mitigates the computational burden of having to train multiple models when approximating the trade-off curve. Crucially, we introduce a novel methodology for quantifying uncertainty in our estimates, thereby providing practitioners with a robust framework for auditing model fairness while avoiding false conclusions due to estimation errors. Our experiments spanning tabular (e.g., Adult), image (CelebA), and language (Jigsaw) datasets underscore that our approach not only reliably quantifies the optimum achievable trade-offs across various data modalities but also helps detect suboptimality in SOTA fairness methods.

Figures (15)

• Figure 1: Accuracy-fairness trade-offs for COMPAS dataset (on held-out data). The black and red curves are obtained using the same optimally trained model evaluated on different splits. The blue curve is obtained using a suboptimally trained model. The green area depicts the range of permissible fairness violations for each accuracy, pink area shows suboptimal accuracy-fairness trade-offs, and blue area shows unlikely-to-be-achieved ones. (Details in Appendix \ref{['subsec:figure_details']})
• Figure 2: Visual illustrations for $\Delta(h)$ (Figure \ref{['fig:delta']}) and our sensitivity analysis procedure (Figure \ref{['fig:sens_anal_illu']}).
• Figure 3: Results on four real-world datasets where $\mathcal{D}_{\textup{cal}}$ is a 10% data split. Here, $\alpha = 0.05$ and we use $|\mathcal{M}|=2$ separately trained models for sensitivity analysis.
• Figure 4: CIs with and without sensitivity analysis for Adult dataset for EO violation.
• Figure 7: CIs obtained by imputing missing senstive attributes using $f_{\mathcal{A}}$ for Adult dataset. Here $n=50$ and $N=2500$.

Theorems & Definitions (14)

• Lemma 3.1: Hoeffding's inequality
• Proposition 3.2
• Proposition 3.3
• Theorem 3.4
• proof : Proof of Lemma \ref{['prop:heoff']}
• proof : Proof of Proposition \ref{['prop:probability-guarantee-ub']}
• proof : Proof of Proposition \ref{['prop:probability-guarantee-lb']}
• Theorem A.3
• proof : Proof of Theorem \ref{['theorem:delta_h_app']}
• Lemma B.1: Hoeffding's inequality