Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Synthetic Data Generation for Intersectional Fairness by Leveraging Hierarchical Group Structure

Gaurav Maheshwari, Aurélien Bellet, Pascal Denis, Mikaela Keller

TL;DR

This work tackles intersectional fairness by recognizing that sensitive attribute groups form a hierarchical structure of intersections and that underrepresented subgroups can be augmented using data from their parent groups. It introduces a modality-agnostic data generation mechanism that synthesizes target-group data by combining parent-group samples and optimizes an $MMD$-based loss, $L_{\mathbf{g},k}(\theta)$, to closely approximate the target distribution $\mathcal{D}_{\mathbf{g}|Y=k}$. Empirical results on four diverse datasets (text and image modalities) show that classifiers trained on augmented data achieve stronger intersectional fairness—without consistently sacrificing overall performance—than several baselines and without exhibiting leveling down. The method is simple, scalable, and adaptable to other distributional divergences, offering a practical approach to improving fairness in real-world deployments across multimodal tasks.

Abstract

In this paper, we introduce a data augmentation approach specifically tailored to enhance intersectional fairness in classification tasks. Our method capitalizes on the hierarchical structure inherent to intersectionality, by viewing groups as intersections of their parent categories. This perspective allows us to augment data for smaller groups by learning a transformation function that combines data from these parent groups. Our empirical analysis, conducted on four diverse datasets including both text and images, reveals that classifiers trained with this data augmentation approach achieve superior intersectional fairness and are more robust to ``leveling down'' when compared to methods optimizing traditional group fairness metrics.

Synthetic Data Generation for Intersectional Fairness by Leveraging Hierarchical Group Structure

TL;DR

This work tackles intersectional fairness by recognizing that sensitive attribute groups form a hierarchical structure of intersections and that underrepresented subgroups can be augmented using data from their parent groups. It introduces a modality-agnostic data generation mechanism that synthesizes target-group data by combining parent-group samples and optimizes an -based loss, , to closely approximate the target distribution . Empirical results on four diverse datasets (text and image modalities) show that classifiers trained on augmented data achieve stronger intersectional fairness—without consistently sacrificing overall performance—than several baselines and without exhibiting leveling down. The method is simple, scalable, and adaptable to other distributional divergences, offering a practical approach to improving fairness in real-world deployments across multimodal tasks.

Abstract

In this paper, we introduce a data augmentation approach specifically tailored to enhance intersectional fairness in classification tasks. Our method capitalizes on the hierarchical structure inherent to intersectionality, by viewing groups as intersections of their parent categories. This perspective allows us to augment data for smaller groups by learning a transformation function that combines data from these parent groups. Our empirical analysis, conducted on four diverse datasets including both text and images, reveals that classifiers trained with this data augmentation approach achieve superior intersectional fairness and are more robust to ``leveling down'' when compared to methods optimizing traditional group fairness metrics.
Paper Structure (35 sections, 15 equations, 4 figures, 5 tables, 1 algorithm)

This paper contains 35 sections, 15 equations, 4 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Problem Statement
  4. Problem Statement:
  5. Approach
  6. Structure of the Data
  7. Data Generation
  8. Parameterization of the Generative Function:
  9. Training the Generative Models:
  10. Training Classifiers on Augmented Data:
  11. Alternative formulations:
  12. Experiments
  13. Datasets:
  14. Methods:
  15. Fairness Metrics:
  16. ...and 20 more sections

Figures (4)

  • Figure 1: Snippet of the hierarchical structure found in intersectional fairness for Twitter Hate Speech Dataset huang-etal-2020-multilingual with 3 sensitive attributes. Here, 'M' stands for Male, 'AA' African American, and 'U45' age under 45 years. The group labeled 'M,AA,U45', represents African American men who are less than 45 years old, and has parent groups 'M,AA', 'M,U45', and 'AA,U45'. For each group, the number of examples is reported. The deeper we go in this hierarchical structure, the smaller the number of examples. Our approach consists in generating additional data for smaller groups by combining data from parent groups.
  • Figure 2: FPR of worst-off group on CelebA (the lower, the better) by varying the number of sensitive axes.
  • Figure 3: $\text{IF}_{0.5}$ comparison between Augmented and Alternate by varying the number of sensitive axes on CelebA. With a smaller number of sensitive axes, Unconstrained and Alternate exhibit comparable performance. However, as the number of sensitive axes increases, Augmented begins to outperform Alternate.
  • Figure 4: Value of $\text{IF}_{\alpha}$ on the test set of various datasets by varying $\alpha \in [0,1]$.