Can Generative Ai-based Data Balancing Mitigate Unfairness Issues in Machine Learning?

Benoît Ronval, Siegfried Nijssen, Ludwig BothmannEWAF 2024 · OpenReview

Abstract

This study investigates the efficacy of Generative AI (GenAI), specifically GANs and LLMs, in balancing datasets to mitigate algorithmic bias stemming from underrepresented protected attributes. By comparing modern GenAI methods (CTGAN, GReaT) against traditional techniques like SMOTE, we find that while balancing generally improves fairness metrics, GenAI does not consistently outperform simpler methods in predictive performance. A persistent trade-off between fairness and accuracy is observed, suggesting that while GenAI offers slight advantages in fairness, it may not be a universal replacement for established resampling techniques.

Introduction

Machine learning models often inherit or exacerbate biases present in their training data. A common source of this algorithmic bias is data imbalance within a "protected attribute" (PA), such as gender or race. Because standard models optimize for overall empirical risk, they are incentivized to fit the majority group closely while neglecting the minority group, leading to unfair automated decisions.

A natural countermeasure is to balance the dataset by acquiring more minority data. Since collecting real-world data is often impossible, researchers turn to artificial data augmentation. While techniques like SMOTE have long been used to balance target variables, this paper investigates whether modern Generative AI (GenAI) can be used to balance protected attributes specifically, and whether these complex models actually yield fairer downstream classifiers than traditional baselines.

Methods

The authors evaluate how different data balancing techniques impact the predictive performance and fairness of downstream classifiers, specifically focusing on Random Forests.

Data Generation Models

  • CTGAN: A Conditional Tabular Generative Adversarial Network designed specifically for tabular data.
  • GReaT: A Large Language Model approach (fine-tuned distil-GPT2) that treats tabular data generation as a text-generation task.
  • SMOTE: A classic, non-generative baseline that synthesizes minority samples by interpolating between existing data points.

Dataset Correction Strategies

  • Real-only: Randomly downsample the majority group to match the minority group.
  • Fake-only: Create a perfectly balanced dataset using only synthetic data.
  • Mixture-full: Keep all original data and generate fake minority samples until the groups are balanced.

Models were evaluated on predictive performance (Accuracy, AUC) and classical fairness metrics: Demographic Parity (DP), Equalized Odds (EO), and Conditional Use Accuracy Equality (CUAE) across the German Credit and Berkeley Admission datasets.

Key Results

Table 1: Comprehensive performance and fairness metrics for German and Berkeley datasets.

Balancing and Fairness

Across the board, balancing the number of majority and minority samples in the training data tends to improve the fairness metrics of the resulting models. On the German dataset, CTGAN achieved a Demographic Parity (DP) of 0.043, compared to SMOTE's 0.062 and the real-only baseline's 0.060. On the Berkeley dataset, CTGAN achieved a DP of 0.159 compared to SMOTE's significantly higher 0.398.

The "Fake-only" Trap and Holdout Testing

Training a classifier exclusively on synthetic data ("fake-only") is generally ineffective. While it might look balanced, models trained this way suffer a drastic drop in predictive performance when evaluated on real-world test data.

Table 2: Results on the German dataset showing performance drops in fake-only scenarios.

As shown in Table 2, when tested on a 20% holdout of real data, the "fake-only" GReaT model's accuracy plummeted to 0.695, compared to 0.760 for the "mixture-full" GReaT model. This underscores that synthetic data cannot entirely replace real observations; real data must remain in the loop (the "mixture-full" approach).

Complexity vs. Superiority

Resource-intensive GenAI models (CTGAN, GReaT) do not clearly outshine the much older, simpler SMOTE algorithm. There is a distinct trade-off: SMOTE tends to win on raw predictive power (e.g., achieving 0.779 Accuracy on German Credit), while GenAI edges out slightly on fairness metrics in specific contexts.

Discussion and Conclusion

While GenAI provides a sophisticated toolset for data augmentation, it does not yet provide a "silver bullet" for the fairness-accuracy trade-off. The study notes that "classical fairness metrics" (like DP and EO) lack deep philosophical justification and are used here primarily as descriptive tools.

Future work will expand experiments to more real-world datasets and conduct thorough simulation studies to test varying degrees of data imbalance and complexity. The authors also intend to evaluate specialized "fair" GenAI models, such as TabFairGAN, which are explicitly designed to generate representative, unbiased tabular data.

Summary updated

BibTeX
@inproceedings{ronval2024can,
  author = {Benoît Ronval and Siegfried Nijssen and Ludwig Bothmann},
  title = {{Can Generative Ai-based Data Balancing Mitigate Unfairness Issues in Machine Learning?}},
  booktitle = {EWAF},
  year = {2024},
  url = {https://openreview.net/forum?id=NaPgBs0TiX},
}
Figure