Fairness-Aware Synthesis

This technique modifies the source training data before the generative model is trained. The goal is to create a 'de-biased' dataset that the generator learns from.

• Reweighting: Adjusts the sampling weights of data points to balance distributions across protected groups.
• Disparate Impact Remover: Applies a linear transformation to non-protected features to make their distributions similar across groups, as defined by the fairlearn library.
• Suppression: Simply removes the protected attribute column from the training data. This is often insufficient as proxies for the attribute may remain in other features.

This method integrates fairness constraints directly into the training objective of the generative model, typically a GAN. An adversarial discriminator is trained to predict the protected attribute from the synthetic data.

• The primary generator aims to create realistic data.
• The adversarial discriminator tries to identify the protected attribute.
• The generator receives a penalty for creating data that allows the discriminator to succeed. This creates a minimax game where the generator learns to produce data from which the protected attribute cannot be inferred, enforcing statistical independence.

This approach applies fairness corrections after synthetic data has been generated. It acts as a filter on the generator's output.

• Resampling: Selects or rejects generated samples to achieve a desired distribution across protected groups (e.g., equal representation).
• Latent Space Interpolation: For models like VAEs, traversing the latent space to find directions that vary non-protected features while holding the protected attribute constant.
• Constraint Satisfaction: Uses optimization to minimally adjust generated feature values to satisfy fairness metrics like demographic parity or equalized odds.

This advanced technique uses Structural Causal Models (SCMs) to generate data that adheres to defined causal relationships, allowing for more nuanced fairness interventions.

• The model encodes causal graphs specifying relationships between protected attributes (A), other features (X), and outcomes (Y).
• Synthesis can be constrained to satisfy counterfactual fairness: "Would this individual have received the same outcome if their protected attribute had been different?"
• This allows for the generation of counterfactual examples essential for auditing model fairness and debugging.

The core training objective of the generative model is modified by adding a fairness regularization term to the standard loss (e.g., adversarial loss for GANs, ELBO for VAEs).

• Correlation Penalty: Adds a term that minimizes the statistical correlation (e.g., using HSIC - Hilbert-Schmidt Independence Criterion) between the protected attribute and other features in the synthetic batch.
• Demographic Parity Loss: Penalizes the generator if the distribution of a key synthetic feature (e.g., loan approval) differs across protected groups.
• This provides a direct, tunable knob to trade off between data fidelity and fairness.

Frames fairness-aware synthesis as a Pareto optimization problem, balancing multiple, often competing, objectives.

Key objectives typically include:

• Realism: Synthetic data should match the real data's distribution (e.g., minimized Wasserstein Distance).
• Utility: Models trained on synthetic data should perform well on real data (TSTR evaluation).
• Fairness: The data should satisfy one or more group fairness metrics.
• Privacy: The data should provide formal guarantees (e.g., differential privacy).

Techniques like Multi-Gradient Descent are used to navigate the trade-off surface and find optimal generator parameters.

Produces synthetic candidate profiles and employee performance data to remove historical hiring biases. Key applications include:

• Training resume screening algorithms on data where gender and ethnicity are decorrelated from skills and experience.
• Simulating diverse workforce scenarios for promotion path analysis.
• Generating balanced datasets for attrition prediction models to ensure fairness across demographic groups, preventing the perpetuation of existing inequities.

Creates synthetic patient cohorts to address underrepresentation in medical datasets, which is critical for precision medicine. This enables:

• Augmenting trials for rare diseases or underrepresented demographic groups (e.g., specific ethnicities, ages).
• Generating counterfactual patient records for causal inference studies on treatment efficacy across groups.
• Developing diagnostic models with robust performance across all patient subpopulations, not just the majority.

Generates synthetic defendant or offender data to audit and retrain recidivism prediction models like COMPAS. The process:

• Identifies and severs spurious correlations between protected attributes (e.g., race) and outcome variables.
• Creates balanced datasets where risk factors are based solely on legally permissible variables.
• Provides a sandbox for developing next-generation assessment tools that meet algorithmic fairness standards before deployment in sensitive, real-world settings.

Synthesizes customer interaction and purchase history data to eliminate bias in personalization engines and customer lifetime value (CLV) models. This application:

• Prevents models from unfairly excluding demographic groups from high-value offers or advertisements.
• Ensures recommendation systems do not reinforce stereotypes (e.g., only showing STEM toys to boys).
• Allows for the safe sharing of realistic customer data for analytics across business units without privacy or discrimination risks.

Creates benchmark synthetic datasets used as a ground truth for algorithmic auditing. Organizations and regulators use this to:

• Test if a deployed model exhibits disparate treatment or disparate impact on synthetically generated protected groups.
• Validate compliance with regulations like the EU AI Act or New York City Local Law 144, which require bias assessments.
• Provide a reproducible, controlled environment for fairness testing without exposing real sensitive data.

A rigorous mathematical framework that provides a formal privacy guarantee for data analysis and generation. It ensures that the output of an algorithm (like a synthetic data generator) does not reveal whether any specific individual's data was included in the input dataset.

• Core Mechanism: Adds calibrated statistical noise to queries or model parameters.
• Relation to Fairness: Often used in conjunction with fairness-aware synthesis to provide both privacy and bias mitigation guarantees, as in the PrivBayes algorithm.
• Key Parameter: Epsilon (ε), which quantifies the privacy loss; a smaller ε offers stronger privacy.

A training-time fairness intervention technique where a primary model (e.g., a classifier) is trained for a task, while an adversarial model is simultaneously trained to predict a protected attribute (like gender) from the primary model's representations. The primary model is optimized to maximize task performance while minimizing the adversary's accuracy.

• Goal: To learn representations that are informative for the main task but invariant to the protected attribute.
• Use in Synthesis: Can be integrated into the training loop of a generative model (like a GAN) to ensure the synthetic data it produces encodes minimal information about sensitive attributes.

A group fairness metric that requires the prediction outcome (or, in synthesis, the distribution of a generated feature) to be independent of membership in a protected group. Also known as demographic parity.

• Formula: P(Ŷ=1 | A=a) = P(Ŷ=1 | A=b) for all groups a, b, where Ŷ is the model's prediction and A is the protected attribute.
• In Synthesis: A fairness-aware generator aims for statistical independence between synthetic features and protected attributes, ensuring synthetic data does not perpetuate disparate impact. For example, the rate of loan approval in synthetic records should be equal across demographic groups, all else being equal.

A causal notion of fairness which states that a model's prediction for an individual should be the same in the actual world and in a counterfactual world where the individual belonged to a different demographic group, keeping all else equal.

• Basis: Relies on a Structural Causal Model (SCM) to model cause-effect relationships.
• In Synthesis: The most rigorous form of fairness-aware synthesis involves generating data from a causal model. The synthetic data should reflect the world as if discriminatory causal pathways (e.g., where race directly influences credit score) did not exist, while preserving other legitimate correlations.

Pre-processing techniques used to adjust training data distributions to mitigate bias before model training or data generation.

• Re-weighting: Assigns different importance weights to training instances based on their group membership and label, effectively creating a balanced loss function.
• Resampling: Oversamples underrepresented groups or undersamples overrepresented groups (e.g., SMOTE for the minority class).
• Relation to Synthesis: These are simpler, distribution-alignment precursors to full fairness-aware synthesis. A synthetic data generator can be explicitly conditioned to produce more samples from underrepresented (group, label) combinations, achieving a balanced synthetic dataset.

The broader field encompassing algorithms, metrics, and processes designed to detect, mitigate, and audit unwanted biases in ML models. It operates across three stages: pre-processing (data), in-processing (algorithm), and post-processing (predictions).

• Key Distinction: While fairness-aware ML focuses on debiasing a predictive model, fairness-aware synthesis focuses on debiasing the training data itself.
• Synergy: Debiased synthetic data serves as a powerful pre-processing tool for training fairer models. Techniques like adversarial debiasing can be applied within both paradigms.