Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fairness Interventions: A Study in AI Explainability

Thomas Souverain, Johnathan Nguyen, Nicolas Meric, Paul Égré

TL;DR

The paper investigates how fairness interventions in AI can be made explainable by examining not only the target fairness criterion but also the variables that constrain its realization. It introduces FairDream, an in-processing, reweighting-based fairness tool that makes its mechanism transparent to lay users and tends to Equalized Odds rather than Demographic Parity in a census income task. Through benchmark comparisons with GridSearch, the work shows that enforcing Demographic Parity can degrade predictive accuracy, whereas FairDream achieves competitive accuracy while reducing disparities under a conditional, true-label–dependent fairness criterion. The authors argue for Equalized Odds as a normative, epistemologically grounded standard in settings where true labels are informative, and advocate for clear explainability about how such corrections operate and respond to data changes.

Abstract

This paper presents a philosophical and experimental study of fairness interventions in AI classification, centered on the explainability of corrective methods. We argue that ensuring fairness requires not only satisfying a target criterion, but also explaining which variables constrain its realization. When corrections are used to mitigate advantage transparently, they must remain sensitive to the distribution of true labels. To illustrate this approach, we built FairDream, a fairness package whose mechanism is made transparent for lay users, increasing the model's weights of errors on disadvantaged groups. While a user may intend to achieve Demographic Parity by the correction method, experiments show that FairDream tends towards Equalized Odds, revealing a conservative bias inherent to the data environment. We clarify the relationship between these fairness criteria, analyze FairDream's reweighting process, and compare its trade-offs with closely related GridSearch models. Finally, we justify the normative preference for Equalized Odds via an epistemological interpretation of the results, using their proximity with Simpson's paradox. The paper thus unites normative, epistemological, and empirical explanations of fairness interventions, to ensure transparency for the users.

Fairness Interventions: A Study in AI Explainability

TL;DR

The paper investigates how fairness interventions in AI can be made explainable by examining not only the target fairness criterion but also the variables that constrain its realization. It introduces FairDream, an in-processing, reweighting-based fairness tool that makes its mechanism transparent to lay users and tends to Equalized Odds rather than Demographic Parity in a census income task. Through benchmark comparisons with GridSearch, the work shows that enforcing Demographic Parity can degrade predictive accuracy, whereas FairDream achieves competitive accuracy while reducing disparities under a conditional, true-label–dependent fairness criterion. The authors argue for Equalized Odds as a normative, epistemologically grounded standard in settings where true labels are informative, and advocate for clear explainability about how such corrections operate and respond to data changes.

Abstract

This paper presents a philosophical and experimental study of fairness interventions in AI classification, centered on the explainability of corrective methods. We argue that ensuring fairness requires not only satisfying a target criterion, but also explaining which variables constrain its realization. When corrections are used to mitigate advantage transparently, they must remain sensitive to the distribution of true labels. To illustrate this approach, we built FairDream, a fairness package whose mechanism is made transparent for lay users, increasing the model's weights of errors on disadvantaged groups. While a user may intend to achieve Demographic Parity by the correction method, experiments show that FairDream tends towards Equalized Odds, revealing a conservative bias inherent to the data environment. We clarify the relationship between these fairness criteria, analyze FairDream's reweighting process, and compare its trade-offs with closely related GridSearch models. Finally, we justify the normative preference for Equalized Odds via an epistemological interpretation of the results, using their proximity with Simpson's paradox. The paper thus unites normative, epistemological, and empirical explanations of fairness interventions, to ensure transparency for the users.
Paper Structure (17 sections, 15 equations, 7 figures, 6 tables)

This paper contains 17 sections, 15 equations, 7 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Age disparities in the evaluation of income
  3. The package FairDream
  4. Demographic Parity vs Equalized Odds
  5. Explaining FairDream's results
  6. An in-processing method
  7. Reweighting for transparent processing
  8. Experimental comparison with GridSearch
  9. Sample Weights vs Optimization under Constraint
  10. Choosing a Fairness Metric conditioned on True Labels
  11. Fairness from Accuracy
  12. Always prefer Equalized Odds?
  13. Conditional dependencies
  14. Conclusion
  15. FairDream's Reweighting
  16. ...and 2 more sections

Figures (7)

  • Figure 1: A sample of the Census dataset
  • Figure 2: Proportion of individuals earning over $50,000 by age. The red and green lines delineate the groups 17-29 and 29-37 between which FairDream detects a disparity.
  • Figure 3: Menu options in FairDream: fairness objective (top) and disparity factor (bottom)
  • Figure 4: Selection of the best corrected model (green), based on a combined measure of statistical performance and fairness.
  • Figure 5: Evaluation on Demographic Parity - Overall Positive Rates with FairDream, GridSearch and Baseline Random Forest Models on Sex
  • ...and 2 more figures