Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ABCFair: an Adaptable Benchmark approach for Comparing Fairness Methods

MaryBeth Defrance, Maarten Buyl, Tijl De Bie

TL;DR

This work applies ABCFair, a benchmark approach which allows adapting to the desiderata of the real-world problem setting, enabling proper comparability between methods for any use case, to a range of pre-, in-, and postprocessing methods on both large-scale, traditional datasets.

Abstract

Numerous methods have been implemented that pursue fairness with respect to sensitive features by mitigating biases in machine learning. Yet, the problem settings that each method tackles vary significantly, including the stage of intervention, the composition of sensitive features, the fairness notion, and the distribution of the output. Even in binary classification, these subtle differences make it highly complicated to benchmark fairness methods, as their performance can strongly depend on exactly how the bias mitigation problem was originally framed. Hence, we introduce ABCFair, a benchmark approach which allows adapting to the desiderata of the real-world problem setting, enabling proper comparability between methods for any use case. We apply ABCFair to a range of pre-, in-, and postprocessing methods on both large-scale, traditional datasets and on a dual label (biased and unbiased) dataset to sidestep the fairness-accuracy trade-off.

ABCFair: an Adaptable Benchmark approach for Comparing Fairness Methods

TL;DR

This work applies ABCFair, a benchmark approach which allows adapting to the desiderata of the real-world problem setting, enabling proper comparability between methods for any use case, to a range of pre-, in-, and postprocessing methods on both large-scale, traditional datasets.

Abstract

Numerous methods have been implemented that pursue fairness with respect to sensitive features by mitigating biases in machine learning. Yet, the problem settings that each method tackles vary significantly, including the stage of intervention, the composition of sensitive features, the fairness notion, and the distribution of the output. Even in binary classification, these subtle differences make it highly complicated to benchmark fairness methods, as their performance can strongly depend on exactly how the bias mitigation problem was originally framed. Hence, we introduce ABCFair, a benchmark approach which allows adapting to the desiderata of the real-world problem setting, enabling proper comparability between methods for any use case. We apply ABCFair to a range of pre-, in-, and postprocessing methods on both large-scale, traditional datasets and on a dual label (biased and unbiased) dataset to sidestep the fairness-accuracy trade-off.
Paper Structure (48 sections, 3 equations, 2 figures, 8 tables)

This paper contains 48 sections, 3 equations, 2 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Contributions.
  3. Related Work.
  4. Comparability Challenges in Fairness Benchmarking
  5. Stage of Intervention
  6. Composition of Sensitive Features
  7. Incompatibility of Fairness Notions
  8. Distribution of the Output
  9. ABCFair
  10. The Dataset Class
  11. Implementation.
  12. Experiment details.
  13. The FairnessMethod class
  14. Implementation.
  15. Experiment details.
  16. ...and 33 more sections

Figures (2)

  • Figure 1: Structural overview of the ABCFair benchmark approach.
  • Figure 2: Fairness-accuracy trade-offs on the SchoolPerformance dataset when trained on biased labels. The top row is evaluated on biased labels and the bottom row on unbiased labels. Each marker is the mean test score over five random seeds, with a confidence ellipse (see Appendix) for one standard deviation.