[Paper Review] DECAF: Generating Fair Synthetic Data Using Causally-Aware Generative Networks

Boris van Breugel, Trent Kyono+2 more|arXiv (Cornell University)|Oct 25, 2021

Ethics and Social Impacts of AISocial Sciences24 references26 citations

TL;DR

DECAF uses a causal GAN to generate fair synthetic tabular data by embedding a structural causal model in the generator and enabling inference-time debiasing via edge removals, with guarantees on downstream fairness and convergence.

ABSTRACT

Machine learning models have been criticized for reflecting unfair biases in the training data. Instead of solving for this by introducing fair learning algorithms directly, we focus on generating fair synthetic data, such that any downstream learner is fair. Generating fair synthetic data from unfair data - while remaining truthful to the underlying data-generating process (DGP) - is non-trivial. In this paper, we introduce DECAF: a GAN-based fair synthetic data generator for tabular data. With DECAF we embed the DGP explicitly as a structural causal model in the input layers of the generator, allowing each variable to be reconstructed conditioned on its causal parents. This procedure enables inference time debiasing, where biased edges can be strategically removed for satisfying user-defined fairness requirements. The DECAF framework is versatile and compatible with several popular definitions of fairness. In our experiments, we show that DECAF successfully removes undesired bias and - in contrast to existing methods - is capable of generating high-quality synthetic data. Furthermore, we provide theoretical guarantees on the generator's convergence and the fairness of downstream models.

Motivation & Objective

• Motivate generating fair synthetic data to ensure downstream models are fair, even when training data are biased.
• Propose a GAN-based framework that embeds a structural causal model to learn conditional data-generating mechanisms.
• Enable inference-time debiasing by removing biased edges in the causal graph to satisfy user-defined fairness definitions.
• Provide theoretical guarantees on generator convergence and fairness of downstream models trained on synthetic data.
• Demonstrate compatibility with multiple fairness definitions and show empirical data utility remains high.

Proposed method

• Model each variable with a dedicated generator G_i conditioned on its causal parents Pa(X_i) and a noise Z_i.
• Train a sequential, DAG-guided conditional GAN where root nodes are generated first and children thereafter (Eq. (2)).
• Use a discriminator D to distinguish generated samples from real data, optimizing a standard GAN objective (Eq. (3)).
• At inference, intervene on the learned causal graph by removing selected edges to satisfy a fairness definition (CF, FTU, DP) using corollaries 1–3.
• Provide a post-processing approach (surrogate do-operations) to generate debiased samples without retraining.
• Prove convergence of the generator distribution to the true data distribution under causally compatible assumptions (Theorem 2).

Experimental results

Research questions

• RQ1 Can a GAN-based generator learn the true causal conditionals of the data while enabling fair synthetic data generation?
• RQ2 How can inference-time interventions on a learned causal graph achieve various fairness notions (FTU, DP, CF) without retraining?
• RQ3 Do synthetic datasets produced by DECAF preserve downstream predictive utility while meeting fairness constraints?
• RQ4 What guarantees exist for the convergence of DECAF and the fairness of downstream models?
• RQ5 How does DECAF perform relative to existing fair data generation approaches across multiple fairness definitions?

Key findings

• DECAF can generate realistic synthetic data while removing undesired bias via inference-time causal edge removal.
• DECAF remains compatible with multiple fairness definitions (FTU, DP, CF) and can tailor the debiasing to the desired notion.
• The method provides theoretical convergence guarantees for the generator distribution to the true data distribution under known DAG compatibility (Theorem 2).
• Empirical results on Adult dataset show DECAF variants achieve higher data utility (precision/recall/AUROC) than baselines under debiasing, with DECAF-DP offering strong DP performance.
• DECAF maintains competitive or superior downstream predictive performance (AUROC) compared to FairGAN and standard GANs while achieving fairness objectives.
• Edge removals at inference-time allow generation of multiple fair datasets without retraining, enabling flexible, task-specific fairness adjustments.