Uncertainty-based Fairness Measures

TL;DR

The paper addresses the limitations of point-based fairness in ML by introducing uncertainty-based fairness measures that capture predictive variance. It leverages Bayesian Neural Networks to decompose predictive uncertainty into epistemic and aleatoric components and defines group- and individual-level fairness metrics such as $\\mathcal{F}_{Epis}$, $\\mathcal{F}_{Alea}$, and $\\mathcal{F}_{Pred}$. The authors demonstrate that these uncertainty-based measures are complementary to traditional fairness metrics and can reveal biases masked by point predictions, using both synthetic datasets and real-world benchmarks (e.g., COMPAS, Adult, D-Vlog). The findings show that uncertainty analyses illuminate sources of bias tied to data scarcity and noise, offering a practical diagnostic tool for fair deployment across diverse settings and highlighting directions for future uncertainty-aware fairness methods.

Abstract

Unfair predictions of machine learning (ML) models impede their broad acceptance in real-world settings. Tackling this arduous challenge first necessitates defining what it means for an ML model to be fair. This has been addressed by the ML community with various measures of fairness that depend on the prediction outcomes of the ML models, either at the group level or the individual level. These fairness measures are limited in that they utilize point predictions, neglecting their variances, or uncertainties, making them susceptible to noise, missingness and shifts in data. In this paper, we first show that an ML model may appear to be fair with existing point-based fairness measures but biased against a demographic group in terms of prediction uncertainties. Then, we introduce new fairness measures based on different types of uncertainties, namely, aleatoric uncertainty and epistemic uncertainty. We demonstrate on many datasets that (i) our uncertainty-based measures are complementary to existing measures of fairness, and (ii) they provide more insights about the underlying issues leading to bias.

Figures (6)

• Figure 1: Existing fairness measures utilize point predictions for quantifying fairness, which ignores the uncertainty (variance) of the predictions (a-b). We fill this gap by using uncertainty instead for measuring fairness (c-d).
• Figure 2: Illustration of Under- or Over-Confidence: An example illustrating under- or over-confidence of an ML model's predictions. The diagram is calculated for a Bayesian NN classifier on the COMPAS Dataset, a dataset frequently used in ML fairness research.
• Figure 3: Two datasets that appear to be fair with point-based measures but unfair in terms of (a) aleatoric uncertainty and (b) epistemic uncertainty. In (c), we see a set where the classifier is fair in terms of uncertainties (both epistemic and aleatoric) but unfair in terms of point-based measures.
• Figure 4: Experiment 3: Point-based (a,b) and uncertainty-based individual fairness (c-f) scores for COMPAS.
• Figure 5: Experiment 3: Point-based (a,b) and uncertainty-based individual fairness (c-f) scores for Adult.

Theorems & Definitions (3)

• Definition 4.1: Uncertainty-Fairness Measure
• Proposition 4.1: Independence of Uncertainty Fairness
• proof