Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Auditing and Mitigating Bias in Gender Classification Algorithms: A Data-Centric Approach

Tadesse K Bahiru, Natnael Tilahun Sinshaw, Teshager Hailemariam Moges, Dheeraj Kumar Singh

TL;DR

The paper tackles bias in gender classification arising from demographic imbalances in training data. It introduces a data-centric four-stage pipeline—dataset audit, targeted repair, fairness-aware training, and comprehensive evaluation—to ground fairness in representativeness, culminating in the BalancedFace dataset. Empirical results show BalancedFace reduces the maximum TPR gap across racial subgroups by over 50% and brings the average Disparate Impact score 63% closer to the ideal of 1.0, with only a minor drop in overall accuracy. This work demonstrates the practical value of data-centric interventions for fair gender classification and provides an openly available resource to advance research in this area.

Abstract

Gender classification systems often inherit and amplify demographic imbalances in their training data. We first audit five widely used gender classification datasets, revealing that all suffer from significant intersectional underrepresentation. To measure the downstream impact of these flaws, we train identical MobileNetV2 classifiers on the two most balanced of these datasets, UTKFace and FairFace. Our fairness evaluation shows that even these models exhibit significant bias, misclassifying female faces at a higher rate than male faces and amplifying existing racial skew. To counter these data-induced biases, we construct BalancedFace, a new public dataset created by blending images from FairFace and UTKFace, supplemented with images from other collections to fill missing demographic gaps. It is engineered to equalize subgroup shares across 189 intersections of age, race, and gender using only real, unedited images. When a standard classifier is trained on BalancedFace, it reduces the maximum True Positive Rate gap across racial subgroups by over 50% and brings the average Disparate Impact score 63% closer to the ideal of 1.0 compared to the next-best dataset, all with a minimal loss of overall accuracy. These results underline the profound value of data-centric interventions and provide an openly available resource for fair gender classification research.

Auditing and Mitigating Bias in Gender Classification Algorithms: A Data-Centric Approach

TL;DR

The paper tackles bias in gender classification arising from demographic imbalances in training data. It introduces a data-centric four-stage pipeline—dataset audit, targeted repair, fairness-aware training, and comprehensive evaluation—to ground fairness in representativeness, culminating in the BalancedFace dataset. Empirical results show BalancedFace reduces the maximum TPR gap across racial subgroups by over 50% and brings the average Disparate Impact score 63% closer to the ideal of 1.0, with only a minor drop in overall accuracy. This work demonstrates the practical value of data-centric interventions for fair gender classification and provides an openly available resource to advance research in this area.

Abstract

Gender classification systems often inherit and amplify demographic imbalances in their training data. We first audit five widely used gender classification datasets, revealing that all suffer from significant intersectional underrepresentation. To measure the downstream impact of these flaws, we train identical MobileNetV2 classifiers on the two most balanced of these datasets, UTKFace and FairFace. Our fairness evaluation shows that even these models exhibit significant bias, misclassifying female faces at a higher rate than male faces and amplifying existing racial skew. To counter these data-induced biases, we construct BalancedFace, a new public dataset created by blending images from FairFace and UTKFace, supplemented with images from other collections to fill missing demographic gaps. It is engineered to equalize subgroup shares across 189 intersections of age, race, and gender using only real, unedited images. When a standard classifier is trained on BalancedFace, it reduces the maximum True Positive Rate gap across racial subgroups by over 50% and brings the average Disparate Impact score 63% closer to the ideal of 1.0 compared to the next-best dataset, all with a minimal loss of overall accuracy. These results underline the profound value of data-centric interventions and provide an openly available resource for fair gender classification research.
Paper Structure (18 sections, 7 equations, 3 figures, 3 tables)

This paper contains 18 sections, 7 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Dataset Audit
  5. Inclusivity
  6. Diversity
  7. Fairness-Aware Training
  8. Model Fairness Evaluation
  9. Equalized Odds
  10. Disparate Impact
  11. Experimental Results
  12. Experimental Setup
  13. Dataset Inclusiveness and Diversity
  14. Model Accuracy and Fairness Trade-off
  15. Fairness Across Racial Groups
  16. ...and 3 more sections

Figures (3)

  • Figure 1: Overview of the proposed method. Each stage produces outputs used by the next, from dataset diagnosis to final fairness evaluation.
  • Figure 2: Comparison of Inclusiveness and Diversity scores across six widely used face datasets. BalancedFace demonstrates markedly higher inclusiveness and especially improved diversity, highlighting its strength in addressing demographic underrepresentation compared to prior datasets.
  • Figure 3: Overall classification accuracy of gender classification models trained on UTKFace, FairFace, and BalancedFace datasets. While UTKFace and FairFace achieve higher raw accuracy, the BalancedFace model achieves more equitable performance across demographic groups with only a minor drop in overall accuracy.