The paper introduces the concept of the fairness Pareto frontier, which characterizes the optimal accuracy-fairness trade-off for a given data distribution and group fairness constraints. This frontier separates aleatoric discrimination, which is inherent in the data, from epistemic discrimination, which is due to algorithmic choices.
The authors first recast the fairness Pareto frontier in terms of the conditional distribution of predicted outcomes given true labels and group attributes. They then use Blackwell's results on comparing statistical experiments to precisely characterize this feasible set of conditional distributions. This allows them to formulate the fairness Pareto frontier as a convex optimization problem.
However, directly solving this optimization problem is challenging, so the authors propose a greedy improvement algorithm that iteratively refines the approximation of the fairness Pareto frontier. They prove convergence guarantees for this algorithm.
The authors apply their framework to benchmark existing group fairness interventions. They find that on standard datasets, state-of-the-art fairness interventions are effective at reducing epistemic discrimination, as their fairness-accuracy curves approach the fairness Pareto frontier. However, when data has disparate missing patterns across groups, aleatoric discrimination increases, diminishing the effectiveness of these fairness interventions.
Overall, the fairness Pareto frontier provides a principled way to separate and quantify different sources of algorithmic discrimination, guiding the development of more effective fairness-enhancing strategies.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Hao Wang,Lux... at arxiv.org 04-17-2024
https://arxiv.org/pdf/2301.11781.pdfDeeper Inquiries