Software Fairness Analysis and Repair via Causal Model-Guided Data Mutation
TL;DR
This work introduces Fairabel, a novel approach that repairs fairness by mutating the training data under the guidance of causal models and multi-objective optimization, and shows that Fairabel reduces ML software bias by 50% on average across three fairness metrics, outperforming the state-of-the-art with a 9% relative improvement.
Abstract
Machine learning (ML) software automates various decision-making processes, significantly enhancing the efficiency of societal operations. However, the widespread adoption of ML software raises growing concerns about fairness, as such software often exhibits biases that disadvantage specific demographic groups or individuals. Since these biases typically stem from correlations between sensitive attributes and label within the training data, disrupting these correlations is a promising direction for fairness repair. Therefore, we introduce Fairabel, a novel approach that repairs fairness by mutating the training data under the guidance of causal models and multi-objective optimization. To evaluate Fairabel, we conduct an extensive empirical comparison against existing approaches across 40 decision-making scenarios. Our experimental results show that Fairabel reduces ML software bias by 50% on average across three fairness metrics, outperforming the state-of-the-art with a 9% relative improvement. Additionally, Fairabel surpasses the state-of-the-art by 1.5% on average across five performance metrics, demonstrating a superior ability to preserve software performance.