FairDiffuseVQVAE Achieves High Fairness in Synthetic Tabular Data

This research introduces FairDiffuseVQVAE, a novel two-stage architecture for generating synthetic tabular data that prioritizes fairness at sampling time. Current state-of-the-art tabular diffusion models excel in data fidelity but lack fairness mechanisms, while existing fairness-aware generators often compromise sample quality or utility. FairDiffuseVQVAE addresses this by separating fidelity and fairness concerns. The first stage involves a vector-quantized autoencoder with a row-level discriminator, focusing purely on data fidelity without fairness terms. The second stage employs a DiffuseVAE-style continuous diffusion refiner that conditions on both the first-stage reconstruction and protected attributes using classifier-free guidance. This design ensures fairness emerges inherently from uniform sampling of protected attributes during inference, rather than through competing loss functions during training. Empirical results on standard datasets demonstrate that FairDiffuseVQVAE achieves significantly higher demographic parity and equalized odds ratios, alongside low correlation error, albeit with a modest trade-off in downstream classifier AUC.