Fairness GAN
Description
Fairness GAN is a pre-processing method that generates a synthetic training dataset in which decisions no longer track a protected attribute such as gender or skin tone. A generative adversarial network (GAN) is a pair of networks trained against each other, one producing fake examples and one trying to tell them from real ones. Here the generator is conditioned on the protected attribute and produces both an example and its decision label, such as hired or rejected (images in the original work, table rows in Xu et al.'s separate FairGAN). An extra network tries to predict the protected attribute from the generated output, and the generator is rewarded for defeating it, so favourable decisions become more similar across groups. The user can aim for demographic parity, meaning equal rates of favourable decisions, or for equality of opportunity, which applies the constraint only to favourable decisions.
Example Use Cases
Fairness
A recruitment technology team holds interview-video stills labelled with past shortlisting decisions that favoured one gender. It trains a Fairness GAN to generate a synthetic image set whose shortlisting labels are similar for women and men, trains the screening model on that set, and compares shortlisting rates by gender on held-out real images before deployment.
A lender's data science team has ten years of loan decisions in which women were approved less often than men on similar incomes. It trains a tabular FairGAN-style generator to produce a synthetic dataset, confirms that the gap in approval rates in the synthetic data is close to zero, and trains its scoring model on that data, testing accuracy and approval rates on real held-out applications.
Limitations
- GAN training is unstable and the fairness loss adds a further pull on the generator. The original authors used spectral normalisation, softened outcome values and added noise to keep training stable, and results still varied across datasets. Expect tuning effort and repeated runs.
- Fairness and accuracy trade off, and the gain is uneven. In the original experiments a classifier trained on debiased CelebA data went from about 20 to about 29 per cent error. Demographic parity improved on three of four datasets but not on Quick, Draw!, and the demographic-parity version made equality of opportunity worse on the soccer data.
- It needs a lot of data. The authors had only 1,501 labelled soccer images, so they borrowed more than 200,000 unlabelled CelebA faces to help the generator, which is possible only when a large related set exists.
- The original method covers a narrow setting: image inputs, one binary protected attribute and a binary outcome, with demographic parity or equality of opportunity but not equalised odds. Continuous protected attributes and non-image data are left as future work, and tabular versions such as FairGAN are separate methods with their own evidence.
- Fairness is judged by training a classifier on the synthetic data and testing it on real data, and image quality is judged mainly by inspecting average faces. The authors report that debiasing visibly changes images in stereotyped ways, such as male faces taking on features coded as feminine, so the synthetic data need checking for distortions that the classifier test would miss.
Resources
Research Papers
Fairness GAN
In this paper, we introduce the Fairness GAN, an approach for generating a dataset that is plausibly similar to a given multimedia dataset, but is more fair with respect to protected attributes in allocative decision making. We propose a novel auxiliary classifier GAN that strives for demographic parity or equality of opportunity and show empirical results on several datasets, including the CelebFaces Attributes (CelebA) dataset, the Quick, Draw!\ dataset, and a dataset of soccer player images and the offenses they were called for. The proposed formulation is well-suited to absorbing unlabeled data; we leverage this to augment the soccer dataset with the much larger CelebA dataset. The methodology tends to improve demographic parity and equality of opportunity while generating plausible images.
FairGAN: Fairness-aware Generative Adversarial Networks
Fairness-aware learning is increasingly important in data mining. Discrimination prevention aims to prevent discrimination in the training data before it is used to conduct predictive analysis. In this paper, we focus on fair data generation that ensures the generated data is discrimination free. Inspired by generative adversarial networks (GAN), we present fairness-aware generative adversarial networks, called FairGAN, which are able to learn a generator producing fair data and also preserving good data utility. Compared with the naive fair data generation models, FairGAN further ensures the classifiers which are trained on generated data can achieve fair classification on real data. Experiments on a real dataset show the effectiveness of FairGAN.
Bt-GAN: Generating Fair Synthetic Healthdata via Bias-transforming Generative Adversarial Networks
Synthetic data generation offers a promising solution to enhance the usefulness of Electronic Healthcare Records (EHR) by generating realistic de-identified data. However, the existing literature primarily focuses on the quality of synthetic health data, neglecting the crucial aspect of fairness in downstream predictions. Consequently, models trained on synthetic EHR have faced criticism for producing biased outcomes in target tasks. These biases can arise from either spurious correlations between features or the failure of models to accurately represent sub-groups. To address these concerns, we present Bias-transforming Generative Adversarial Networks (Bt-GAN), a GAN-based synthetic data generator specifically designed for the healthcare domain. In order to tackle spurious correlations (i), we propose an information-constrained Data Generation Process (DGP) that enables the generator to learn a fair deterministic transformation based on a well-defined notion of algorithmic fairness. To overcome the challenge of capturing exact sub-group representations (ii), we incentivize the generator to preserve sub-group densities through score-based weighted sampling. This approach compels the generator to learn from underrepresented regions of the data manifold. To evaluate the effectiveness of our proposed method, we conduct extensive experiments using the Medical Information Mart for Intensive Care (MIMIC-III) database. Our results demonstrate that Bt-GAN achieves state-of-the-art accuracy while significantly improving fairness and minimizing bias amplification. Furthermore, we perform an in-depth explainability analysis to provide additional evidence supporting the validity of our study. In conclusion, our research introduces a novel and professional approach to addressing the limitations of synthetic data generation in the healthcare domain. By incorporating fairness considerations and leveraging advanced techniques such as GANs, we pave the way for more reliable and unbiased predictions in healthcare applications.