Abstract
Automated gender classification systems often exhibit significant demographic bias, particularly against non-Western populations. This paper introduces "pseudo-balancing," a semi-supervised learning (SSL) strategy that leverages unlabeled, diverse datasets to mitigate these biases. By enforcing demographic parity during the pseudo-label selection process, the method significantly reduces performance gaps—most notably narrowing the gender accuracy gap in East Asian populations from over 49% to just 5.01%—without requiring expensive manual annotations.
Introduction
Automated gender classification systems frequently suffer from "demographic blindness." While they may achieve high accuracy on Western populations, performance often plummets when encountering Black or East Asian faces. This disparity typically stems from biased training datasets and algorithms that inadvertently amplify these biases during learning.
While curating perfectly balanced datasets is a common solution, it is prohibitively expensive. Semi-supervised learning (SSL) offers a cost-effective alternative by learning from a small amount of labeled data and a large pool of unlabeled data. However, standard SSL techniques like FixMatch often suffer from the "Matthew effect": the model generates pseudo-labels based on its own biased internal logic, effectively reinforcing its initial prejudices.
The Pseudo-Balancing Method
The authors introduce pseudo-balancing, a lightweight modification to iterative self-training pipelines (specifically FixMatch and FlexMatch). In a standard SSL pipeline, a model predicts labels for unlabeled images; if confident, those images are treated as ground truth for the next round of training. Pseudo-balancing adds a crucial constraint to break the bias feedback loop.
 Source, p. 4
The process involves three key steps:
- Partitioning: High-confidence predictions are divided into gender groups.
- Subsampling: These groups are manually subsampled to ensure an equal number of "pseudo-male" and "pseudo-female" samples are used in the next training iteration.
- Dynamic Adjustment: This works alongside confidence-based selection to prevent the model from being overwhelmed by a majority class, maintaining class equilibrium without requiring ground-truth labels for the secondary dataset.
Experimental Setup and Datasets
The study utilizes three primary datasets to evaluate bias mitigation:
- Kaggle Gender Classification Dataset: Used to train the initial biased baseline model.
- FairFace (FF): A demographically balanced dataset used without labels to test if its diversity alone could debias a model.
- All-Age-Faces (AAF): The primary evaluation benchmark, consisting predominantly of East Asian faces where Western-trained models typically struggle.
Key Results and Discussion
Pseudo-balancing consistently improves fairness while often increasing overall accuracy. The most striking success was observed in the East Asian subgroup, where the baseline model's massive 49% gender gap was narrowed to just 5.01%.
Key performance metrics include:
- Overall Accuracy: Reached 79.81%, a 6.53% improvement over the baseline.
- Gender Gap Reduction: The overall gender accuracy gap was reduced by 44.17%.
- Selection Rate (SR): The method achieved a 93.80% Selection Rate, indicating successful bias mitigation (where >80% is the target threshold).
Conclusion and Future Work
Pseudo-balancing demonstrates that you don't need labels to fix a biased model; you just need a diverse, unlabeled "mirror" to reflect a more balanced world. By forcing the model to "slow down" and wait for confident samples from under-represented groups, the method breaks the positive feedback loop of bias.
However, the method relies on the initial model's ability to generate at least some correct pseudo-labels. Future work is needed to address more complex intersectional biases involving age, facial expression, and cultural context, as well as scenarios with extreme initial data skew.
Summary updated