Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Intersectionality Problem for Algorithmic Fairness

Johannes Himmelreich, Arbie Hsu, Kristian Lum, Ellen Veomett

TL;DR

This paper elucidates the problem of intersectionality in algorithmic fairness, develops desiderata to clarify the challenges underlying the problem and guide the search for potential solutions, and develops desiderata to clarify the challenges underlying the problem.

Abstract

A yet unmet challenge in algorithmic fairness is the problem of intersectionality, that is, achieving fairness across the intersection of multiple groups -- and verifying that such fairness has been attained. Because intersectional groups tend to be small, verifying whether a model is fair raises statistical as well as moral-methodological challenges. This paper (1) elucidates the problem of intersectionality in algorithmic fairness, (2) develops desiderata to clarify the challenges underlying the problem and guide the search for potential solutions, (3) illustrates the desiderata and potential solutions by sketching a proposal using simple hypothesis testing, and (4) evaluates, partly empirically, this proposal against the proposed desiderata.

The Intersectionality Problem for Algorithmic Fairness

TL;DR

This paper elucidates the problem of intersectionality in algorithmic fairness, develops desiderata to clarify the challenges underlying the problem and guide the search for potential solutions, and develops desiderata to clarify the challenges underlying the problem.

Abstract

A yet unmet challenge in algorithmic fairness is the problem of intersectionality, that is, achieving fairness across the intersection of multiple groups -- and verifying that such fairness has been attained. Because intersectional groups tend to be small, verifying whether a model is fair raises statistical as well as moral-methodological challenges. This paper (1) elucidates the problem of intersectionality in algorithmic fairness, (2) develops desiderata to clarify the challenges underlying the problem and guide the search for potential solutions, (3) illustrates the desiderata and potential solutions by sketching a proposal using simple hypothesis testing, and (4) evaluates, partly empirically, this proposal against the proposed desiderata.

Paper Structure

This paper contains 35 sections, 13 equations, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Algorithmic Fairness
  4. The Problem of Intersectionality
  5. Existing Work
  6. Kearns et al.
  7. Foulds et al. and Morina et al.
  8. Molina and Loiseau
  9. Cherian and Candès
  10. Khan et al. and Agrawal et al.
  11. Desiderata
  12. Minimal Justice
  13. Consistent Conceptualization
  14. Incentive Compatibility
  15. Two Alternative Metrics
  16. ...and 20 more sections

Figures (5)

  • Figure 1: Plots of accuracy $m(G)$, optimist's metric $c_1^g$, and pessimist's metric $c_2^g$ of critical subgroups $G$ for each dataset. The $x$-axis corresponds to the percentage of the critical subgroup that is kept. Legend lists the dataset name.
  • Figure 2: Showing the relationship between $m$ (metric), $n$ (number in subgroup), and $c$ (edge of confidence interval). Hues in \ref{['fig:2dlower']} shows the values of $c$ in the pessimist's model. Hues in \ref{['fig:2dupper_with_min']} shows the values of $c$ in the optimist's model, but with an upper limit of 1 (since no proportion can be larger than 1). Hues in \ref{['fig:2dupper']} shows the values of $c$ in the optimist's model, without limiting the value at 1 (so that we can see more easily where it is very difficult to reject the optimist's hypothesis that the model is fair).
  • Figure 3: Plots of $m(G), c_1^G$, and $c_2^G$ of critical subgroups $G$ for each dataset. Here we subsampled the entire dataset, and the $x$-axis corresponds to the percentage of the entire dataset that is kept. Legend lists the dataset name.
  • Figure 4: Values of expression \ref{['eqn:Kearns_metric']} on the adult, meps20, titanic, and bank datasets. Horizontal axis is the percent of the subgroup, vertical axis is unfairness (i.e., the value of expression \ref{['eqn:Kearns_metric']}).
  • Figure 5: Values of expression \ref{['eqn:Kearns_metric']} on the adult, meps20, titanic, and bank datasets. Horizontal axis is the percent of the entire dataset kept, vertical axis is unfairness (i.e., the value of expression \ref{['eqn:Kearns_metric']}).