Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Benefits and Risks of Transductive Approaches for AI Fairness

Muhammed Razzak, Andreas Kirsch, Yarin Gal

TL;DR

Transductive learning uses a holdout set to guide training, but the holdout's sensitive-group composition can bias fairness outcomes. The authors study two transductive methods, RHOS-Loss and FairGen, on CIFAR100-20 and CelebA to assess how holdout balance affects discriminative accuracy, fairness metrics, and generative quality. They find that imbalanced holdouts exacerbate disparities while balanced holdouts can mitigate biases across both discriminative and generative tasks, underscoring the practical impact of holdout design. The results emphasize the need for diverse, representative holdout sets and motivate developing strategies to construct or adapt holdouts to deployment distributions for robust, fair transductive learning.

Abstract

Recently, transductive learning methods, which leverage holdout sets during training, have gained popularity for their potential to improve speed, accuracy, and fairness in machine learning models. Despite this, the composition of the holdout set itself, particularly the balance of sensitive sub-groups, has been largely overlooked. Our experiments on CIFAR and CelebA datasets show that compositional changes in the holdout set can substantially influence fairness metrics. Imbalanced holdout sets exacerbate existing disparities, while balanced holdouts can mitigate issues introduced by imbalanced training data. These findings underline the necessity of constructing holdout sets that are both diverse and representative.

The Benefits and Risks of Transductive Approaches for AI Fairness

TL;DR

Transductive learning uses a holdout set to guide training, but the holdout's sensitive-group composition can bias fairness outcomes. The authors study two transductive methods, RHOS-Loss and FairGen, on CIFAR100-20 and CelebA to assess how holdout balance affects discriminative accuracy, fairness metrics, and generative quality. They find that imbalanced holdouts exacerbate disparities while balanced holdouts can mitigate biases across both discriminative and generative tasks, underscoring the practical impact of holdout design. The results emphasize the need for diverse, representative holdout sets and motivate developing strategies to construct or adapt holdouts to deployment distributions for robust, fair transductive learning.

Abstract

Recently, transductive learning methods, which leverage holdout sets during training, have gained popularity for their potential to improve speed, accuracy, and fairness in machine learning models. Despite this, the composition of the holdout set itself, particularly the balance of sensitive sub-groups, has been largely overlooked. Our experiments on CIFAR and CelebA datasets show that compositional changes in the holdout set can substantially influence fairness metrics. Imbalanced holdout sets exacerbate existing disparities, while balanced holdouts can mitigate issues introduced by imbalanced training data. These findings underline the necessity of constructing holdout sets that are both diverse and representative.
Paper Structure (21 sections, 3 equations, 3 figures, 4 tables)

This paper contains 21 sections, 3 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Outline
  3. Related Work
  4. Fairness Reweighting Methods
  5. Fairness in Generative Models
  6. Background
  7. Transductive Training Methods
  8. Reducible Holdout Loss (RHO-Loss)
  9. FairGen
  10. Fairness Metrics
  11. Discriminative Setting
  12. Generative Setting
  13. Empirical Study
  14. Datasets
  15. Methods
  16. ...and 6 more sections

Figures (3)

  • Figure 1: Illustration of the online batch setting set-up as used in mindermann2021.
  • Figure 2: True Positive Rate of different classes and protected attributes on CIFAR20 on the balanced and highly imbalanced holdout sets.
  • Figure 3: The dataset proportions and importance weights for the CelebA dataset used in the generative modeling setting. Minority class (males) are downweighted and Majority class upweighted as a result of the imbalance in the reference dataset.