Source-linked AI summary
FairGAN: Fairness-aware Generative Adversarial Networks
Depeng Xu, Shuhan Yuan, Lu Zhang, Xintao Wu
TL;DR
Discriminatory patterns in real data can persist in synthetic data and affect downstream classification. FairGAN uses a generator with two discriminators to produce fair synthetic data while preserving utility, and is reported to support fair classification on real data relative to naive approaches.
Problem
Synthetic data can inherit discrimination from real data, so fair data generation must address both generated-data fairness and downstream classification fairness.
Method
FairGAN conditions generation on the protected attribute and uses one discriminator for real-data similarity and another to remove protected-attribute correlations.
Results
FairGAN is reported to generate fair data with good utility and to enable classifiers trained on generated data to achieve fair classification on real data.
Takeaways & Limitations
FairGAN provides a data-generation approach intended to address disparate treatment and disparate impact while retaining useful synthetic data.
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
The paper motivates fair data generation as a way to preserve utility while preventing discrimination in generated datasets and downstream classification. It frames FairGAN as addressing limitations of naive approaches that randomize protected attributes without removing disparate impact.
- Historical-data-based machine learning can produce biased performance and discriminatory decisions against protected groups.
- Synthetic data can inherit discrimination from real data because high similarity preserves discriminatory patterns used in downstream predictive analysis.
- Fair data generation must preserve data utility, data fairness, classification utility, and classification fairness simultaneously.
- Naive generation keeps unprotected attributes and decisions close to real data while randomly assigning protected attributes, removing disparate treatment but not disparate impact.
- FairGAN generates revised unprotected attributes and decisions conditioned on the protected attribute, targeting independence from that attribute.
2 RELATED WORK
Prior work addresses discrimination through data preprocessing, fair representations, predictive constraints, and adversarial training. FairGAN instead generates discrimination-free data while targeting utility and applicability across numerical and categorical attributes.
- Preprocessing methods modify labels, weights, samples, or non-label attributes to reduce discrimination while preserving data utility.
- Fairness-aware predictive models balance predictive accuracy against discriminatory predictions using statistical-parity or individual-fairness constraints.
- Adversarial-training studies learn fair representations or optimize objectives for group-fairness notions such as demographic parity, equalized odds, and equal opportunity.
- FairGAN generates fair datasets rather than modifying existing data, aiming to remove disparate treatment and disparate impact while preserving generation utility.
- Unlike methods limited to numerical attributes, FairGAN is described as applicable to both numerical and categorical data and can generate additional training data.
3 PRELIMINARY
The preliminaries define dataset and classifier fairness through statistical parity, risk differences, and ϵ-fairness, then introduce GANs as adversarial distribution-matching models. They also explain autoencoder-based generation for discrete data.
- Fairness and Discrimination: Dataset statistical parity requires equal positive-decision rates across protected groups, measured by the risk difference disc(D).
- Fairness and Discrimination: Classifier statistical parity measures the difference in positive prediction rates between protected groups through disc(η).
- Fairness and Discrimination: ϵ-fairness evaluates whether the protected attribute is predictable from unprotected attributes, using balanced error rate as the prediction-error measure.
- Generative Adversarial Network: GANs use a generator and discriminator in an adversarial game, with the generator learning PG to match the real distribution Pdata.
- Generative Adversarial Network: For discrete data, medGAN combines a generator with an autoencoder decoder that converts generated salient representations into synthetic samples.
4 FAIRGAN
FairGAN conditions generation on the protected attribute and uses two discriminators: one matches real joint data, while the other removes protected-attribute information. This design targets fair synthetic data and fair classification on real data.
- Model: FairGAN generates protected-attribute-conditioned pairs (ˆx, ˆy) and forms joint samples from PG(x,y,s) = PG(x,y|s)PG(s), with PG(s) = Pdata(s).
- Model: Discriminator D1 distinguishes real joint data from generated data, encouraging the generator to learn the real joint distribution.
- Model: Discriminator D2 distinguishes generated samples conditioned on s = 1 from those conditioned on s = 0, encouraging equal conditional distributions.
- Model: The model uses λ to trade off data-generation utility against fairness during optimization.
- Model: The generator fools D2 so generated samples cannot predict the protected attribute, removing correlations associated with disparate impact.
- Training: FairGAN pretrains an autoencoder and then fine-tunes its decoder to remove discrimination information while optimizing the generator.
V (GDec, D1, D2)
FairGAN combines fidelity to real joint data with similarity between protected-group conditional distributions, while naive variants satisfy only narrower fairness goals. Theoretical analysis and toy experiments show why this distinction matters for fair classification on real data.
- Theoretical analysis: The FairGAN objective balances Jensen-Shannon divergence from real joint data against divergence between the two generated conditional distributions.Theorem 1 gives C(GDec) = −(2 + λ) log 4 + 2 · JSD(Pdata(x,y,s)||PG(x,y,s)) + 2λ · JSD(PG(x,y|s = 1)||PG(x,y|s = 0)).
- Theoretical analysis: The two fairness objectives cannot be achieved simultaneously in general, but convexity allows both JSD terms to converge to a global optimum.The optimum is expressed as −(2 + λ) log 4 + ∆, where ∆ is the minimum value reached by the two JSD terms.
- NaïveFairGAN-II: NaïveFairGAN-II enforces equal conditional synthetic distributions and statistical parity, yet likewise preserves disparate impact because it does not preserve real conditional distributions.Its ideal condition is PG(x,y|s = 1) = PG(x,y|s = 0) = (Pdata(x,y|s = 1) + Pdata(x,y|s = 0))/2.
- NaïveFairGAN-I: NaïveFairGAN-I preserves PG(x) close to Pdata(x) and makes protected attributes independent, but it leaves disparate impact through correlations with the real protected attribute.Its classifier therefore cannot achieve fair prediction when tested on real data.
- Toy experiment: FairGAN makes generated protected-group conditionals resemble the real conditionals while making generated unprotected attributes unable to predict the real protected attribute.In the toy example, similar conditional means are reported as removing disparate impact, unlike both naive models.
5 EXPERIMENTS
The experiments evaluate FairGAN for fair data generation and fair classification on real data, using utility and fairness metrics against GAN-based baselines. FairGAN removes disparate impact while retaining good utility and enabling fairer real-data classification.
- Experimental setup: FairGAN is evaluated against regular GAN, NaïveFairGAN-I, and NaïveFairGAN-II for fair data generation and fair classification.The experiments use the UCI Adult income dataset and synthetic datasets generated by each model.
- Fair data generation: Figure 4 compares conditional probabilities given s = 1 and s = 0, with the diagonal representing identical distributions and ideal fairness.Each dot represents one attribute.
- Fair data generation: FairGAN removes disparate impact by generating unprotected attributes that do not correlate with the real protected attribute.SYN4-FairGAN achieves a BER of 0.3862±0.0036 when predicting real s with a linear SVM.
- Fair data generation: FairGAN retains good data utility after removing disparate impact, although its distributions are not necessarily as close to the real data as regular GAN.Utility is assessed using dimension-wise probabilities and Euclidean distances over joint and conditional probabilities.
- Fair classification: Classifiers trained on NaïveFairGAN data remain significantly discriminatory on real data, whereas classifiers trained on FairGAN data have small risk differences on real data.This distinction arises because NaïveFairGAN preserves correlations between generated unprotected attributes and the real protected attribute.
- Fair classification: FairGAN classifiers preserve a good utility–fairness balance, with SYN4-FairGAN accuracies only slightly below SYN1-GAN accuracies on real data.The small utility loss is attributed to modifying unprotected attributes to remove disparate impact.
- Parameter analysis: Increasing λ lowers generated-data and classifier risk differences while classifier accuracy remains relatively steady with a slight decrease.The standard deviations of accuracies across λ values are smaller than 1%.
6 CONCLUSIONS AND FUTURE WORK
The paper concludes that FairGAN generates fair data free from disparate treatment and disparate impact while retaining high data utility. It identifies equalized odds and equal opportunity on real data as future extensions.
- Conclusion: FairGAN generates data free from disparate treatment and disparate impact in terms of the real protected attribute while retaining high data utility.Its generator is conditioned on the protected attribute and trained against separate real-or-fake and group-discrimination discriminators.
- Conclusion: Classifiers trained on FairGAN-generated data are not subject to discrimination when making decisions on real data.The paper reports experimental evidence supporting this outcome.
- Future work: Future work will extend FairGAN so classifiers achieve equalized odds or equal opportunity on real data in addition to statistical parity.The paper distinguishes this planned work from its current focus.
APPENDIX
The appendix introduces the theoretical analysis of NaïveFairGAN-II and identifies the optimal discriminators used in its value-function analysis.
- Theoretical analysis: The appendix analyzes NaïveFairGAN-II theoretically.The proof procedure is described as a straightforward extension of the proof in Proposition 1.
- Optimal discriminators: For fixed G, the analysis derives optimal discriminators D1 and D2 for the value function V (GDec, D1, D2).These discriminator solutions are then used to reformulate the minimax game.
V (GDec, D1, D2)
The value-function analysis rewrites the NaïveFairGAN-II minimax objective using expectations under real and generated conditional distributions. Jensen–Shannon divergence then characterizes its global minimum.
- Value function: The reformulated objective combines expectations over real and generated samples conditioned on s = 1 and s = 0.The terms include log ratios involving real and generated conditional distributions.
- Divergence form: C′(G) = −4 log 4 + 2 · JSD(PG(x,y|s = 1)||PG(x,y|s = 0)).The objective is rewritten as a Jensen–Shannon divergence expression.
- Distribution matching: The divergence form compares the sum of real conditional distributions with the sum of generated conditional distributions.This comparison appears in the reformulated value-function expression.
- Global minimum: The global minimum occurs when PG(x,y|s = 1) = PG(x,y|s = 0) = [Pdata(x,y|s = 1) + Pdata(x,y|s = 0)]/2.This follows because Jensen–Shannon divergence is non-negative and zero only for equal distributions.