Learning Antidote Data to Individual Unfairness

TL;DR

This paper introduces a novel approach to mitigate individual unfairness in machine learning by generating on-manifold antidote data that aligns with data distribution, improving fairness with minimal utility loss.

Contribution

It proposes learning and generating antidote data that follow data distribution to address individual unfairness, overcoming limitations of previous adversarial perturbation methods.

Findings

Effective reduction of individual unfairness on tabular datasets.

Minimal or no loss in predictive utility compared to baselines.

Compatible with pre-processing and in-processing fairness paradigms.

Abstract

Fairness is essential for machine learning systems deployed in high-stake applications. Among all fairness notions, individual fairness, deriving from a consensus that `similar individuals should be treated similarly,' is a vital notion to describe fair treatment for individual cases. Previous studies typically characterize individual fairness as a prediction-invariant problem when perturbing sensitive attributes on samples, and solve it by Distributionally Robust Optimization (DRO) paradigm. However, such adversarial perturbations along a direction covering sensitive information used in DRO do not consider the inherent feature correlations or innate data constraints, therefore could mislead the model to optimize at off-manifold and unrealistic samples. In light of this drawback, in this paper, we propose to learn and generate antidote data that approximately follows the data distribution to remedy individual unfairness. These generated on-manifold antidote data can be used through a generic optimization procedure along with original training data, resulting in a pure pre-processing approach to individual unfairness, or can also fit well with the in-processing DRO paradigm. Through extensive experiments on multiple tabular datasets, we demonstrate our method resists individual unfairness at a minimal or zero cost to predictive utility compared to baselines.