FairReweighing: Density Estimation-Based Reweighing Framework for Improving Separation in Fair Regression
Xiaoyin Xi, Zhe Yu
TL;DR
This paper introduces FairReweighing, a density estimation-based pre-processing method that enhances fairness in regression models by reducing separation violations, applicable to both binary and continuous sensitive attributes.
Contribution
It proposes a novel FairReweighing algorithm inspired by classification reweighing, extending fairness measures to regression with theoretical guarantees and superior empirical performance.
Findings
Outperforms existing fairness methods in regression tasks
Guarantees separation under data independence assumptions
Effective on both synthetic and real-world datasets
Abstract
There has been a prevalence of applying AI software in both high-stakes public-sector and industrial contexts. However, the lack of transparency has raised concerns about whether these data-informed AI software decisions secure fairness against people of all racial, gender, or age groups. Despite extensive research on emerging fairness-aware AI software, up to now most efforts to solve this issue have been dedicated to binary classification tasks. Fairness in regression is relatively underexplored. In this work, we adopted a mutual information-based metric to assess separation violations. The metric is also extended so that it can be directly applied to both classification and regression problems with both binary and continuous sensitive attributes. Inspired by the Reweighing algorithm in fair classification, we proposed a FairReweighing pre-processing algorithm based on density…
Peer Reviews
Decision·Submitted to ICLR 2024
The paper includes a comprehensive summary of related work. The ideas of the paper are clearly presented. The author proposes a universal preprocessing framework, which is a training framework based on density estimation. By adjusting the impact of each data item through weight allocation to achieve fairness, classification and regression models can be effectively trained. This article extends the fairness issue in classification problems to regression problems based on previous studies.
Lack of analysis on the robustness and stability of the algorithm: The paper did not analyze the robustness and stability of the algorithm. In practical applications, algorithms need to be able to handle various uncertainties and noise while maintaining stable performance. The analysis of the robustness and stability of algorithms can provide a more comprehensive evaluation.
Regression fairness has received only negligible attention compared to classification fairness, and any progress on this important topic is highly welcome. The presented approach is very straightforward - that being a good thing -, easy to implement, and very general and widely applicable to all kinds of models and training schemes. The paper is well-written and easy to read.
As outlined above, I consider the topic important and the solution proposed by the authors straightforward and generally promising. I do, however, sadly see quite a few significant weaknesses in the present manuscript. Firstly, I am honestly quite confused about what it is that the authors actually implemented, and how that matches the textual description. Sections 1 and 2 suggest strongly that the authors implement a scheme to achieve separation, i.e., $\hat{Y} \perp A \mid Y$. However, the we
- The motivation is clear, and the criterion of separation in regression problem receives relatively less attention in the fairness literature. - The presented material is clear and easy-to-follow.
1. This paper uses KDE to deal with difficulties in using finite samples to achieve fairness, but does not provide any analysis for the proposed procedures for auditing and bias mitigation. - What is the convergence rate of $\hat r_\mathrm{sep}$ to $r_\mathrm{sep}$ (eqs. 4 and 7)? - Similarly, what is the convergence rate of $\widehat W(A,Y)$ to $W(A,Y)$ (eq. 9), when estimated and inferred from finite samples using KDE/radius neighbors/binning? 2. Although an intuition is provided for the p
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsEthics and Social Impacts of AI · Adversarial Robustness in Machine Learning · Explainable Artificial Intelligence (XAI)