Fairness-Aware Synthesis

Ensures that the probability of a favorable outcome is independent of protected attributes such as race or gender. The generator is constrained so that the synthetic distribution exhibits equal base rates across subgroups. This is achieved through adversarial debiasing—where a critic network penalizes the generator for producing biased outputs—or by resampling and reweighting latent representations to balance outcome distributions. The result is a dataset where decision-making models trained upon it will not disproportionately favor one demographic over another.

A stricter fairness criterion requiring that true positive and false positive rates are equal across protected groups. In synthesis, this is enforced by conditioning the generative model on sensitive attributes and applying post-processing corrections to the learned distribution. Techniques include counterfactual data augmentation, where synthetic records are generated with flipped sensitive attributes while holding other features constant, and constrained optimization of the generator's loss function. This prevents the model from using protected attributes as proxies for outcomes.

Automated workflows that quantify representation disparities in real data before synthesis and validate fairness metrics in the synthetic output. These pipelines typically involve:

• Disparate impact ratio calculation across subgroups
• Kolmogorov-Smirnov tests for distributional divergence
• Causal graph analysis to identify discriminatory pathways
• Iterative retraining of the generator with fairness penalties Tools like Fairlearn and AI Fairness 360 integrate with synthetic data libraries to provide standardized fairness audits.

Moves beyond statistical parity by modeling the underlying causal structure of the data. The generator learns a structural causal model (SCM) that distinguishes between legitimate causal relationships and spurious correlations driven by bias. By intervening on the causal graph—specifically, severing edges from protected attributes to outcomes—the synthesis process generates data that is fair under counterfactual fairness definitions. This approach prevents the model from exploiting proxy variables that are causally downstream from sensitive attributes.

The inherent tension between achieving perfect fairness and preserving predictive accuracy. Fairness-aware synthesis must navigate this Pareto frontier by:

• Applying Lagrangian relaxation to treat fairness as a constrained optimization problem
• Using multi-objective evolutionary algorithms to discover optimal trade-off points
• Reporting fairness-utility curves that allow stakeholders to select an acceptable operating point Excessive fairness constraints can degrade the synthetic data's fidelity for downstream tasks, requiring careful calibration.

Addresses bias across combinations of protected attributes rather than treating each in isolation. A synthetic dataset may appear fair for gender and race separately but still disadvantage women of a specific racial group. Intersectional fairness-aware synthesis uses subgroup-specific fairness metrics and hierarchical conditioning in the generative model to ensure equitable representation across all intersectional strata. This requires sufficient sample sizes in the real data for each subgroup or specialized few-shot generation techniques.

A fairness criterion requiring that a model's positive prediction rate is equal across all protected subgroups. In synthetic data generation, demographic parity is enforced by conditioning the generative process on sensitive attributes to ensure the output distribution is independent of group membership.

• Formula: P(ŷ=1|A=a) = P(ŷ=1|A=b) for all groups a, b
• Use case: Loan approval models where acceptance rates must be balanced across demographics
• Limitation: Does not account for legitimate group-level differences in base rates

A stricter fairness metric requiring equal true positive and false positive rates across protected groups. Equalized odds ensures that the model's errors are distributed evenly, preventing one group from bearing a disproportionate burden of misclassification.

• True Positive Rate Parity: P(ŷ=1|Y=1, A=a) = P(ŷ=1|Y=1, A=b)
• False Positive Rate Parity: P(ŷ=1|Y=0, A=a) = P(ŷ=1|Y=0, A=b)
• Synthetic data application: Rejection sampling or constrained decoding to match error rates across subgroups

A causal definition of fairness stating that a prediction for an individual should remain the same in a counterfactual world where their protected attribute is changed. Counterfactual fairness requires a structural causal model to generate synthetic individuals with flipped sensitive attributes.

• Requires: Explicit causal graph of the data-generating process
• Synthesis approach: Intervene on the sensitive attribute node during generation
• Strength: Captures downstream effects of discrimination through causal pathways

A preprocessing bias mitigation strategy that assigns instance weights to training examples based on group membership and label distribution. In fairness-aware synthesis, reweighing is used to oversample underrepresented protected group combinations before training the generative model.

• Weight formula: W(a,y) = P_exp(A=a ∧ Y=y) / P_obs(A=a ∧ Y=y)
• Effect: Creates a balanced synthetic dataset without discarding any real records
• Compatible with: CTGAN, TVAE, and other tabular generative models

Techniques that embed fairness objectives directly into the loss function or architecture of Generative Adversarial Networks. FairGAN and similar architectures add a discriminator that attempts to predict protected attributes from generated samples, penalizing the generator when group membership is recoverable.

• Adversarial debiasing: Generator minimizes a fairness adversary's accuracy
• Lagrangian optimization: Treats fairness as a constrained optimization problem
• Result: Synthetic data where sensitive attributes are independent of other features