🤖 machine learning

FairDiffuseVQVAE: Sampling-Time Fairness in Tabular Diffusion via Conditional Refinement of Vector-Quantized Latents

FairDiffuseVQVAE introduces a novel two-stage architecture that decouples data fidelity from fairness by using a vector-quantized autoencoder followed by a conditional diffusion refiner, achieving superior demographic parity and equalized odds on tabular datasets without compromising sample quality through explicit fairness penalties during training.

Original authors: Nitish Nagesh, Mahdi Bagheri, Amir M. Rahmani

Published 2026-08-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nitish Nagesh, Mahdi Bagheri, Amir M. Rahmani

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a chef trying to recreate a famous, complex dish for a massive banquet. You have a recipe (the data) that tells you exactly how the original dish tastes, but there's a catch: the original recipe was written by a biased cook who always added extra salt to the soup if the guest was wearing a red hat, and extra pepper if they wore a blue one. If you just follow the recipe exactly, your new banquet will taste just as unfair. This is the world of synthetic data: creating fake but realistic datasets to train computers without exposing real people's private information. The goal is to make the fake data look so real that a computer learns from it, but without copying the unfair "red hat" bias.

To do this, scientists use two main tools. First, there are diffusion models, which are like a sculptor who starts with a block of noisy, shapeless clay and slowly chips away the noise until a perfect statue emerges. Second, there are vector-quantized autoencoders, which act like a translator that turns a messy, long sentence into a short, efficient code (like a secret handshake) and then back into a sentence again. The big question researchers are asking is: How do we make sure the final statue doesn't accidentally favor red hats over blue ones, without ruining the statue's shape or making it look fake?

Enter FairDiffuseVQVAE, a new two-stage recipe for creating fair data. The authors suggest that instead of trying to force fairness into the sculptor's hands while they are still chipping away the clay (which often ruins the statue's shape), we should let the sculptor focus purely on making a beautiful, accurate statue first. Then, at the very end, we use a special "fairness filter" during the final polishing stage.

Here is how their method works, step-by-step:

Stage 1: The Honest Translator
First, the system builds a "translator" (a vector-quantized autoencoder). This part looks at the real data and learns to compress it into a compact, efficient code and then rebuild it perfectly. Crucially, this stage is told to ignore the "hat color" (the protected attribute) entirely. It just wants to make the best possible copy of the data, with no fairness penalties. It's like a photographer taking a high-resolution picture of a crowd without worrying about who is wearing what; they just want the picture to be sharp and true to life.

Stage 2: The Fair Refiner
Once the translator has made a rough draft of the data, a second tool steps in: a diffusion refiner. This is where the magic happens. This tool is trained to take the rough draft and polish it, but it has a special instruction: it must be able to generate data for any hat color, not just the ones that were common in the original photo.

The key trick is Classifier-Free Guidance. Imagine the refiner is an artist who can paint a scene based on a description. Usually, they paint based on what they saw in the real world. But here, the artist is trained to sometimes ignore the description of the hat color (during training). Then, when it's time to create the final fake data, the artist is given a specific instruction: "Paint a scene where exactly half the people wear red hats and half wear blue hats." Because the artist was trained to understand how to paint for any hat color, they can follow this new instruction perfectly. They don't need to be punished for being unfair; they just need to be told to be fair at the moment of creation.

The Results: A Trade-off
The paper tested this on real-world datasets like loan applications and criminal justice records. The results were striking. By using this "fairness at the end" approach, the new method achieved a Demographic Parity Ratio of 0.702, which is a 47% improvement over the previous best method (FairTabDDPM). It also hit an Equalized Odds Ratio of 0.686, a massive 100% jump.

However, the paper is honest about the cost. To get this level of fairness, the system had to sacrifice a bit of "utility." The accuracy of a classifier trained on this new data dropped by about 15 AUC points compared to the most accurate (but unfair) models. The authors explain this makes sense: if you force the data to be perfectly fair, the computer can no longer use the old, biased shortcuts to make predictions, so its performance on real-world tasks might dip slightly. It's a deliberate trade: you get a fairer system, but you might lose a little bit of raw predictive power.

What It Doesn't Do
The paper also clarifies what this method is not. It doesn't require a complex map of cause-and-effect relationships (like a causal graph) to work, which makes it easier to use than some other methods. However, it doesn't offer a mathematical proof that it will always be fair in every possible scenario; it works because of how the sampling is done, not because of a theoretical guarantee. Also, on very small datasets, the system sometimes "memorized" the training data too well, which is a risk for privacy, though it worked well on larger datasets.

In short, FairDiffuseVQVAE shows that we don't have to choose between a realistic model and a fair one by fighting over the training process. Instead, we can build a realistic model first, and then simply "tune the knob" at the end to ensure the final output treats everyone equally. It's a clever, practical way to bake fairness into the recipe without ruining the taste.

Technical Summary: FairDiffuseVQVAE

Problem Statement

Synthetic tabular data is critical for privacy-preserving sharing, data augmentation, and mitigating downstream classifier bias. While state-of-the-art diffusion models like TabDDPM and TabSyn achieve high distributional fidelity, they inherit statistical biases from training data without mechanisms for fairness. Conversely, existing fairness-aware generators (e.g., DECAF, FairTGAN, FairTabDDPM) typically impose explicit fairness penalties during training. This approach often forces the model to balance reconstruction against adversarial or auxiliary losses, resulting in modest fairness gains at the cost of significant degradation in sample quality (density estimation error) or downstream utility.

Methodology: FairDiffuseVQVAE

The authors propose FairDiffuseVQVAE, a two-stage architecture that decouples fidelity from fairness. Instead of competing loss terms, fairness is enforced as a property of the sampling distribution via conditional refinement.

Stage 1: Fair Vector-Quantized Autoencoder

This stage focuses exclusively on learning a faithful, compact representation of mixed-type tabular data without any fairness constraints.

Architecture: A VQGAN-style autoencoder consisting of an encoder ( $E_\phi$ ), a discrete codebook ( $Z$ ), a decoder ( $G_\theta$ ), and a row-level discriminator ( $D_\xi$ ).
Process: The encoder maps a row $\bar{x}$ to latent tokens $z_e$ , which are quantized to the nearest codebook entry $z_q$ . The decoder reconstructs the row conditioned on the protected attribute $s$ , while the encoder remains unconditioned.
Losses: The model minimizes reconstruction loss (MSE for continuous, Cross-Entropy for categorical), codebook commitment loss, and an adversarial GAN loss. Crucially, no fairness term is included in this stage.

Stage 2: Conditional Input-Space Diffusion Refiner

This stage refines the Stage-1 reconstruction in the input space to enforce fairness.

Architecture: A continuous diffusion refiner (DiffuseVAE-style) using an EDM (Elucidating the Design Space) formulation. It employs a 5-layer MLP denoiser ( $D_\psi$ ).
Conditioning: The denoiser is conditioned on three inputs:
1. The frozen Stage-1 reconstruction ( $\hat{\bar{x}}_{VQ}$ ).
2. The noise level $\sigma$ .
3. The protected attribute $s$ .
Training: The model is trained with Classifier-Free Guidance (CFG). During training, the conditioning signal $s$ is stochastically dropped (with probability $p_{cfg}=0.3$ ) to learn both conditional $p(\bar{x}|s)$ and unconditional $p(\bar{x})$ distributions.
Sampling & Fairness Enforcement:
1. The protected attribute $s$ is drawn from a uniform distribution (e.g., $Bernoulli(0.5)$) rather than the biased data marginal.
2. Random codebook indices are decoded to form an initial guess.
3. A 50-step Heun EDM sampler refines this guess conditioned on the uniform $s$ and the Stage-1 reconstruction.
4. The guidance scale $w_{cfg}$ acts as a tunable knob to control the strength of the conditional manifold.

By sampling $s$ uniformly and conditioning the diffusion process on it, the synthetic joint distribution $p_{synth}(\bar{x}, s) = Unif(s) \cdot p(\bar{x}|s)$ satisfies Demographic Parity by construction.

Key Contributions

Novel Architecture: Introduction of a two-stage framework combining a VQGAN-style autoencoder with an input-space DiffuseVAE refiner, specifically designed for fair tabular synthesis.
Sampling-Time Fairness: Demonstration that fairness can be achieved without explicit loss penalties by leveraging classifier-free guidance and uniform sampling of the protected attribute during inference.
State-of-the-Art Performance: The method achieves the highest mean Demographic Parity Ratio (DPR) and Equalized Odds Ratio (EOR) on the Yang et al. (2025) benchmark, significantly outperforming prior fairness-aware baselines.
High Fidelity: The architecture attains the lowest mean pair-wise column correlation error among all published methods, surpassing even the high-fidelity TabSyn.
Reproducibility: A reproducible pipeline integrated into the TabSyn repository, covering eight datasets and reporting a comprehensive suite of metrics (density, correlation, AUC, DPR, EOR, DCR).

Experimental Results

Evaluated on the Yang et al. (2025) benchmark (Adult, Bank, COMPAS) and the TabSyn benchmark:

Fairness Metrics: FairDiffuseVQVAE achieved a mean Demographic Parity Ratio (DPR) of 0.702 (a +47% improvement over FairTabDDPM) and a mean Equalized Odds Ratio (EOR) of 0.686 (a +100% improvement).
Quality Metrics: It achieved the lowest mean pair-wise correlation error (0.034), outperforming TabSyn (0.041).
Utility Trade-off: The method explicitly trades downstream utility for fairness. The mean TSTR (Train on Synthetic, Test on Real) AUC was 0.708, approximately 15 points lower than the strongest deep generative baselines (e.g., TabSyn at 0.860). The authors attribute this to the decoupling of features from the protected attribute, preventing downstream classifiers from exploiting bias.
Privacy: Distance to Closest Record (DCR) scores were near ideal (0.5) for Adult and Bank but elevated (0.81) for the smaller COMPAS dataset, indicating partial memorization in small-data regimes.

Significance and Claims

The paper claims that FairDiffuseVQVAE offers a practical alternative to causally-grounded methods (like DECAF) that require explicit causal graph specifications. By decoupling fidelity (Stage 1) from fairness (Stage 2), the architecture allows the model to optimize reconstruction without competing with fairness constraints during training.

The authors emphasize that the utility-fairness trade-off is explicit and quantifiable. They argue that for fairness-aware synthesis, it is the "correct behavior" for a downstream classifier trained on synthetic data to lose accuracy if that accuracy was derived from biased correlations in the real data. The method provides a tunable mechanism (via CFG guidance) to control this trade-off without retraining.

The paper concludes that while the approach is empirical and lacks theoretical guarantees (unlike causal methods), it provides strong empirical results and a flexible framework for sampling-time control of generative behavior. The authors caution that synthetic data should not be treated as a standalone fairness solution and recommend combining their method with qualitative audits and classifier-side interventions.