Pith. sign in

REVIEW 3 major objections 5 minor 57 references

This paper claims that sampling the protected attribute uniformly at inference time makes the synthetic distribution satisfy demographic parity by construction, and that a two-stage VQ autoencoder plus diffusion refiner delivers the best fa

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 16:45 UTC pith:T6AM7XAF

load-bearing objection Novel architecture, honest reporting, but the 'by construction' fairness claim doesn't hold and the baseline comparison is unmatched. the 3 major comments →

arxiv 2607.28945 v1 pith:T6AM7XAF submitted 2026-07-31 cs.LG

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

Nitish Nagesh , Mahdi Bagheri , Amir M. Rahmani This is my paper
classification cs.LG
keywords FairDiffuseVQVAEtabular data synthesisdemographic parityequalized oddsvector-quantized autoencoderdiffusion modelsclassifier-free guidancesynthetic data fairness
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper introduces FairDiffuseVQVAE, a two-stage generator for tabular data that tries to separate the task of learning realistic rows from the task of making the generated data fair. The central claim is that fairness can be moved entirely to sampling time: if the protected attribute is drawn uniformly and the generator is conditioned on that attribute, the synthetic marginal satisfies demographic parity by construction, so no fairness penalty competes with reconstruction during training. Empirically, on the Adult, Bank, and COMPAS fairness benchmark, the paper reports the highest mean Demographic Parity Ratio (0.702) and Equalized Odds Ratio (0.686) among compared methods, with the lowest mean pair-correlation error (0.034), while giving up about 15 points of downstream classifier accuracy (TSTR AUC). A sympathetic reader would care because fair tabular synthesis typically couples fairness losses to training, creating a fidelity-fairness trade-off; this design proposes a tunable, retraining-free control instead.

Core claim

The central discovery is that fairness can be enforced as a property of the sampling distribution rather than a training objective. The paper builds a two-stage architecture: Stage 1 is a vector-quantized autoencoder—a row is mapped to discrete codebook tokens and decoded back, with a row-level discriminator, and no fairness term in its loss. Stage 2 is an input-space diffusion refiner that learns the conditional distribution of a row given the protected attribute, using classifier-free guidance. At sampling, the protected attribute is drawn from a Bernoulli(0.5) distribution and random codebook tokens are decoded and refined; because the marginals for the two protected groups are forced to

What carries the argument

The load-bearing object is the conditional sampling distribution p_synth(¯x,s)=Unif(s)p(¯x|s), which makes demographic parity an identity rather than an optimized goal. It is realized by two components. Stage 1, a vector-quantized autoencoder (an autoencoder that represents each row as discrete codebook tokens), learns to reconstruct mixed-type rows with a row-level discriminator and no fairness term; its encoder never sees the protected attribute. Stage 2 trains an EDM denoiser in the original input space, conditioned on the frozen Stage-1 reconstruction and the protected attribute, with classifier-free-guidance dropout. At sampling, the protected attribute is drawn uniformly, random codebo

Load-bearing premise

The load-bearing premise for the headline margins is that published baseline numbers are directly comparable to the paper's own runs, yet the baselines were not re-run under this paper's protocol (which selects the best of two classifiers per seed and uses three seeds with high variance); if that comparability fails, the reported +47% DPR and +100% EOR gains could shrink materially.

What would settle it

Re-run every baseline through the paper's exact pipeline—same three seeds, same per-seed best-of-two-classifier selection, same TSTR evaluation script—and recompute DPR and EOR. If the margins over FairTabDDPM fall below roughly +47% and +100%, the headline fairness claim fails. Separately, generate a large synthetic sample and measure the protected-attribute marginal: under the paper's mechanism it should be statistically indistinguishable from 0.5/0.5, and a synthetic-only DPR computed on the generated data should approach 1; any large deviation indicates the uniform-sampling construction is

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the claim holds, fair tabular data can be generated from an existing trained model without retraining: changing the protected-attribute prior or the guidance weight at inference time moves the fairness-utility trade-off.
  • Because Stage 1 has no fairness term, improvements in reconstruction fidelity should transfer directly to better conditional density estimation, with fairness behavior controlled independently by the sampling procedure.
  • The explicit ~15-point AUC cost quantifies how much demographic parity costs in downstream utility; practitioners who need both should combine this generator with in-processing fair classifiers rather than expecting fairness and accuracy to come free.
  • The same sampling-time mechanism should extend to multi-valued protected attributes and to conditional generation tasks beyond fairness, since it only requires a conditional model and a choice of attribute prior.
  • On small datasets, the input-space refiner's elevated DCR scores indicate the architecture's privacy behavior is not uniform across dataset sizes; latent-space refinement is the natural next step noted by the authors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reported margins (+47% DPR, +100% EOR) rest on baseline numbers taken as published; a matched re-evaluation with the same per-seed classifier selection could shrink the apparent gains substantially, since the paper's own variance (σ_DPR up to 0.24) dwarfs published baseline variances.
  • The 'by construction' guarantee applies to the synthetic marginal, not to the downstream classifier's behavior on real data; a fairer synthetic marginal does not automatically yield a fair classifier when the real test distribution has different base rates, and the paper itself cautions that fairness metrics on synthetic-trained classifiers do not measure decision fairness.
  • A testable extension would be to feed the model a deliberately skewed protected-attribute prior (e.g., Bernoulli(0.9)) and check whether the resulting synthetic distribution tracks the requested group ratio; this would separate the uniform-sampling mechanism from the conditional model's ability to represent both groups.
  • The architecture's decoupling suggests a recipe for other fairness notions: swap the protected-attribute prior to match a target parity criterion (e.g., equalized odds requires conditioning on the outcome as well), without retraining the denoiser.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. FairDiffuseVQVAE is a two-stage tabular data generator that combines a vector-quantized autoencoder (Stage 1) with an input-space EDM diffusion refiner (Stage 2) conditioned on the protected attribute via classifier-free guidance. The paper claims that at sampling time, drawing the protected attribute uniformly from Bernoulli(0.5) 'enforces demographic parity by construction,' and reports large improvements in Demographic Parity Ratio (mean 0.702, +47%) and Equalized Odds Ratio (mean 0.686, +100%) over FairTabDDPM on the Yang et al. (2025) benchmark, with the lowest pair-correlation error among published methods and an explicit ~15-point TSTR-AUC sacrifice. The paper is candid in Section 8 about high seed variance, COMPAS memorization, and the absence of a theoretical guarantee, and it ships a reproducible pipeline integrated into TabSyn.

Significance. The architectural idea of decoupling fidelity from fairness—training a high-fidelity generator without fairness penalties and performing fairness control at sampling time—is potentially valuable, and the release of a reproducible pipeline is a strength. However, the central formal claim that uniform s-sampling enforces demographic parity by construction is incorrect as stated, and the experimental comparison uses a favorable classifier-selection protocol (per-seed best-of-two) against baselines taken from the literature. If these issues are corrected by reframing the contribution as an empirical heuristic and providing matched baseline evaluations, the method could still be a useful empirical contribution to fair tabular synthesis. As it stands, the headline SOTA claims are not adequately supported.

major comments (3)
  1. [§4.3, Eq. (1)] The claim 'uniform sampling of the protected attribute at inference time enforces demographic parity by construction' (Abstract; also §1 and §4.3) is not supported by the paper's math. DPR and EOR (Eqs. 1–2) are evaluated by training a classifier on synthetic data and testing on real data, so the relevant quantities are Pr(Yhat=1|A=a) under the real conditional distributions P(X|A=a). Uniform s in the synthetic training set does not change P(X|A) in the real test distribution, nor does it make X conditionally independent of A. The argument in §4.3 writes the gap as |E_{p(·|0)}[g]−E_{p(·|1)}[g]|, which is just the definition of the gap, not a proof that it vanishes; it vanishes only if p(x|s=0)=p(x|s=1), which is exactly the original bias. Moreover, the paper uses w_cfg=1 throughout (§5.4), so the 'guidance mechanism' that is said to pull the conditional distributions together is inactive
  2. [§5.2, Table 4] The headline margins (+47% DPR, +100% EOR) compare 'Ours'—for which the best of two classifiers (LR vs. MLP) is selected per seed and per metric (Table 4 caption)—against baselines that were not re-run under the same protocol; §5.2 states 'We report all baseline numbers as published.' Given the high seed-to-seed variance of the proposed method (σ_DPR up to 0.24 in Table 4, versus σ≈0.005–0.07 for published baselines as acknowledged in §8), per-seed best-of-two selection can bias the reported means upward. Without a matched evaluation—same classifier-selection rule, same seeds, same downstream evaluation script—the claim of state-of-the-art fairness is not established. Please re-run the baselines under your own protocol, or report the results for both classifiers separately without per-seed selection, and quantify the effect of the selection rule.
  3. [§5.4, Figure 1] Every reported result uses w_cfg=1, i.e., no classifier-free guidance at sampling time. The paper repeatedly attributes the fairness gains to the guidance mechanism ('the conditional distributions ... are pulled closer together by the guidance mechanism,' §4.3; CFG as a 'tunable fairness knob,' §4.3). With w_cfg=1, sampling reduces to standard conditional sampling, so the experiments provide no evidence that CFG contributes to the observed DPR/EOR improvements. At minimum, please include an ablation varying w_cfg (e.g., w_cfg = 1, 2, 4) on at least one dataset, or revise the text to attribute the improvement to conditional generation rather than to guidance strength.
minor comments (5)
  1. [§6.3] The text says 'Adult and Bank show DCR within 12% and 8% of the ideal respectively,' but Adult DCR in Table 3 is 0.357, which is 28.6% below the ideal 0.5, while Bank at 0.538 is about 7.6% above. The Adult claim appears to be a numerical error.
  2. [§5.3, Table 4] Please clarify whether the per-seed best-of-two classifier selection was applied to every baseline. If not, the comparison is not apples-to-apples; this should be described explicitly in the limitations or appendix.
  3. [Abstract and throughout] The phrase 'enforces demographic parity by construction' is contradicted by Section 8 ('No theoretical guarantee'). Please align the wording with the actual scope of the claim.
  4. [Table 1] Ours is not the best method on Adult DPR (0.533 vs. FairTGAN 0.554) or Adult EOR (0.594 vs. FairTGAN 0.697). The headline mean comparisons should be accompanied by a per-dataset discussion so readers can see the heterogeneous effect.
  5. [General] There are typographical issues in the rendered text, e.g., 'FairDiffuseVQV AE' with an odd space, and 'F airDiffuseVQV AE' in the abstract and Figure 1. These should be cleaned up.

Circularity Check

1 steps flagged

Headline metrics are externally measured, but the 'demographic parity by construction' framing is self-definitional; no self-citation or fitted-input circularity.

specific steps
  1. self definitional [Abstract; §1 'Our perspective'; §4.3 'Why this enforces fairness']
    "uniform sampling of the protected attribute at inference time enforces demographic parity by construction (Abstract); §4.3: 'the marginal synthetic distribution p_synth(¯x, s) = Unif(s)·p(¯x|s) satisfies demographic parity by construction'."

    The input to the construction is s∼Bernoulli(0.5), so p_synth(s) = Unif(s). If 'demographic parity' means balance of the protected-attribute marginal in the generated dataset, the claimed property is identical to the sampling input — an identity, not a derived result. If 'demographic parity' instead means the DPR/EOR of Eqs. 1–2 — a downstream classifier's group-wise rates on the real test set — then uniform s does not imply it, since the classifier is evaluated on biased real conditionals p(x|s); §8 concedes 'no theoretical guarantee' and possible counterexamples. The 'by construction' statement therefore either collapses into its own definitional input or is disconnected from the headline metric.

full rationale

The headline numbers (DPR 0.702, EOR 0.686, pair-corr 0.034, TSTR AUC 0.708) are measurements on held-out real test data against external published benchmarks, not quantities derived from the paper's own fitted constants; there is no fitted parameter relabeled as a prediction, and no self-citation chain (the authors cite no prior work of their own). The central architecture borrows VQGAN, EDM, DiffuseVAE and classifier-free guidance from external, independently established literature, so no ansatz is smuggled in via self-citation. The one definitional reduction is the 'demographic parity by construction' claim: read as balance of s in the synthetic marginal it is a tautology equivalent to the sampling input; read as the Eq. 1–2 DPR/EOR measured on real data it does not follow, as §8 explicitly concedes ('no theoretical guarantee', 'counterexamples may exist'). That overreach in framing is the basis for the score of 2; the empirical fairness results themselves are independently measured and not circular. The comparison-protocol asymmetry (published baseline numbers vs. per-seed best-classifier selection for Ours) is a comparability/correctness risk, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central empirical claims depend on a list of design choices (uniform protected sampling, Bank age binarization, w_cfg=1, per-seed classifier selection) and on inherited assumptions about CFG validity, the train/sample match of the Stage-1 prior, and the TSTR protocol's fidelity as a fairness proxy. No new entities are invented. The heaviest unexamined premise is the train/sample gap: Stage-2 is trained on encoder reconstructions but samples from uniform random codebook decodes.

free parameters (5)
  • Uniform protected-attribute sampling at inference (Bernoulli(0.5)) = p(s)=0.5/0.5
    The entire fairness mechanism: changes the synthetic marginal of s by construction, but per the paper's own gap formula it does not by itself equalize p(x|0) and p(x|1), so the measured DPR/EOR gains rest on this design choice plus the downstream classifiers.
  • Bank age binarization threshold = age < 25 (Section 5.1)
    Creates a ~3% minority class on Bank (stated in Section 7); Bank contributes the largest per-dataset gains (DPR 0.687 vs 0.337 and EOR 0.697 vs 0.334 over FairTabDDPM). The headline means depend on this preprocessing.
  • Classifier-free guidance weight w_cfg = 1.0 (all experiments)
    At w=1, Eq. 11 reduces to plain conditional sampling; the advertised 'tunable fairness knob' is never exercised, yet the paper's causal story leans on CFG separating conditional from unconditional predictions.
  • Per-seed best-of-two-classifier selection = LR or 100-unit MLP, chosen per metric per seed (Table 4 caption)
    Applies only to Ours; inflates DPR/EOR/AUC and interacts with the high variance to make reported means optimistic relative to baselines whose numbers came from their own protocols.
  • EDM/CFG and architecture hyperparameters = P_mean=−1.2, P_std=1.2, σ_data=0.5, 50 Heun steps, ρ=7, p_cfg=0.3, K=1024, 8×32 tokens, widths 256/512, β=0.25, λ_adv=0.
    Hand-chosen in Section 5.4 within standard ranges; no sensitivity analysis, so their contribution to the reported trade-off curve is unquantified.
axioms (5)
  • domain assumption Classifier-free guidance with dropout p_cfg=0.3 produces a valid conditional model p(x|s) that the sampler can query at uniform s
    Standard Ho & Salimans (2022) mechanism, but its validity at extreme imbalance (Bank minority ~3%, Section 7) is untested; Section 7 concedes the conditional model must extrapolate.
  • ad hoc to paper Stage-2 denoising conditioned on encoder-produced reconstructions transfers to conditioning on random-code decodes at sampling time
    Section 4.3: training conditions D_ψ on x̂_VQ = G_θ(z_q, s) from the encoder; sampling decodes uniformly random codebook indices {k_i} ~ Unif into the initial guess. The train/sample mismatch in the conditioning input is unanalyzed and plausibly contributes to the 13x worse density error vs TabSyn (Table 2: 0.194 vs 0.015).
  • domain assumption DPR/EOR measured by TSTR (train on synthetic, test on real) reflects the synthetic distribution's fairness property
    Inherited from Yang et al. (2025) and defined in Section 2.4; Section 9 warns the metrics measure the downstream classifier's fairness on the real test distribution, not decision-fairness about real individuals.
  • domain assumption The DPR/EOR results are insensitive to the unspecified input-feature choice for the downstream classifier
    Section 2.4 defines Ŷ without specifying whether the protected attribute is a classifier feature; the interpretation of the DPR numbers (feature-level vs allocation fairness) depends on this choice, which is never described.
  • domain assumption The EDM preconditioned denoiser trained under Eq. 10, sampled with 50 Heun steps (σ_min=0.002, σ_max=80, ρ=7), yields samples from the learned conditional distribution
    Karras et al. (2022) framework; the paper reports no convergence analysis or step-count ablation for this setting.

pith-pipeline@v1.3.0-daily-deepseek · 14829 in / 26434 out tokens · 240151 ms · 2026-08-03T16:45:40.274350+00:00 · methodology

0 comments
read the original abstract

Synthetic tabular data is increasingly used in privacy-preserving data sharing, data augmentation, and to mitigate downstream classifier bias. State-of-the-art tabular diffusion models such as TabDDPM and TabSyn achieve excellent distributional fidelity but offer no mechanism for fairness; conversely, fairness-aware tabular generators (DECAF, FairTGAN, FairTabDDPM) impose explicit fairness penalties at training time, yielding modest fairness gains at substantial cost to either sample quality or downstream utility. We introduce FairDiffuseVQVAE, a two-stage architecture that decouples fidelity from fairness: a vector-quantized autoencoder with a row-level discriminator (Stage~1, no fairness terms) is followed by a DiffuseVAE-style continuous diffusion refiner that conditions on both the Stage-1 reconstruction and the protected attribute via classifier-free guidance (Stage~2). Fairness emerges as a property of the sampling distribution -- uniform sampling of the protected attribute at inference time enforces demographic parity by construction, rather than from competing loss terms. On the Adult, Bank and COMPAS datasets, FairDiffuseVQVAE achieves the highest mean Demographic Parity Ratio ($0.702$, $+47\%$ over FairTabDDPM) and Equalized Odds Ratio ($0.686$, $+100\%$). It also attains the lowest mean pair-wise correlation error ($0.034$) of any published method, while explicitly trading $\sim$$15$ AUC points for these fairness gains.

Figures

Figures reproduced from arXiv: 2607.28945 by Amir M. Rahmani, Mahdi Bagheri, Nitish Nagesh.

Figure 1
Figure 1. Figure 1: Overview of FairDiffuseVQVAE. A two-stage architecture for fair mixed-type tabular syn￾thesis. Stage 1 trains a vector-quantized autoencoder: an encoder Eϕ maps a row ¯x to latent tokens ze, which are quantized against a learned codebook Z and decoded by Gθ into a reconstruction xˆ¯VQ, with a row-level discriminator Dξ providing an adversarial signal. The decoder is conditioned on the protected at￾tribute … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

57 extracted references · 44 linked inside Pith

  1. [1]

    Imposing fairness constraints in synthetic data generation

    Mahed Abroshan, Andrew Elliott, and Mohammad Mahdi Khalili. Imposing fairness constraints in synthetic data generation. In Proceedings of the 27th International Conference on Artificial Intelligence and Statistics (AISTATS), volume 238 of PMLR, 2024

    2024

  2. [2]

    A reductions approach to fair classification

    Alekh Agarwal, Alina Beygelzimer, Miroslav Dud \' k, John Langford, and Hanna Wallach. A reductions approach to fair classification. In Proceedings of the 35th International Conference on Machine Learning (ICML), 2018. arXiv:1803.02453

    Pith/arXiv arXiv 2018

  3. [3]

    Alaa, Boris van Breugel, Evgeny S

    Ahmed M. Alaa, Boris van Breugel, Evgeny S. Saveliev, and Mihaela van der Schaar. How faithful is your synthetic data? sample-level metrics for evaluating and auditing generative models. In Proceedings of the 39th International Conference on Machine Learning (ICML), pages 290--306, 2022. arXiv:2102.08921

    Pith/arXiv arXiv 2022

  4. [4]

    Machine bias: There's software used across the country to predict future criminals

    Julia Angwin, Jeff Larson, Surya Mattu, and Lauren Kirchner. Machine bias: There's software used across the country to predict future criminals. and it's biased against blacks, 2016. ProPublica, May 23, 2016

    2016

  5. [5]

    Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg

    Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. In Advances in Neural Information Processing Systems (NeurIPS), volume 34, 2021. arXiv:2107.03006

    Pith/arXiv arXiv 2021

  6. [6]

    Adult (census income) data set, 1996

    Barry Becker and Ronny Kohavi. Adult (census income) data set, 1996. UCI Machine Learning Repository

    1996

  7. [7]

    Rachel K. E. Bellamy, Kuntal Dey, Michael Hind, Samuel C. Hoffman, Stephanie Houde, Kalapriya Kannan, Pranay Lohia, Jacquelyn Martino, Sameep Mehta, Aleksandra Mojsilovi \'c , Seema Nagar, Karthikeyan Natesan Ramamurthy, John Richards, Diptikalyan Saha, Prasanna Sattigeri, Moninder Singh, Kush R. Varshney, and Yunfeng Zhang. AI fairness 360: An extensible...

    Pith/arXiv arXiv 2019

  8. [8]

    Estimating or propagating gradients through stochastic neurons for conditional computation

    Yoshua Bengio, Nicholas L \'e onard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013

    Pith/arXiv arXiv 2013

  9. [9]

    Erickson, Isabelle Guyon, and Kristin P

    Karan Bhanot, Miao Qi, John S. Erickson, Isabelle Guyon, and Kristin P. Bennett. The problem of fairness in synthetic healthcare data. Entropy, 23 0 (9): 0 1165, 2021

    2021

  10. [10]

    Deep neural networks and tabular data: A survey

    Vadim Borisov, Tobias Leemann, Kathrin Se ler, Johannes Haug, Martin Pawelczyk, and Gjergji Kasneci. Deep neural networks and tabular data: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2022. arXiv:2110.01889

    Pith/arXiv arXiv 2022

  11. [11]

    Language models are realistic tabular data generators

    Vadim Borisov, Kathrin Se ler, Tobias Leemann, Martin Pawelczyk, and Gjergji Kasneci. Language models are realistic tabular data generators. In Proceedings of the 11th International Conference on Learning Representations (ICLR), 2023. arXiv:2210.06280. Introduces GReaT

    Pith/arXiv arXiv 2023

  12. [12]

    Fairness in machine learning: A survey

    Simon Caton and Christian Haas. Fairness in machine learning: A survey. ACM Computing Surveys, 56 0 (7): 0 1--38, 2024. Article 166, arXiv:2010.04053

    Pith/arXiv arXiv 2024

  13. [13]

    Path-specific counterfactual fairness

    Silvia Chiappa. Path-specific counterfactual fairness. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 7801--7808, 2019

    2019

  14. [14]

    Weis, Kevin Swersky, Toniann Pitassi, and Richard Zemel

    Elliot Creager, David Madras, J \"o rn-Henrik Jacobsen, Marissa A. Weis, Kevin Swersky, Toniann Pitassi, and Richard Zemel. Flexibly fair representation learning by disentanglement. In Proceedings of the 36th International Conference on Machine Learning (ICML), 2019. arXiv:1906.02589

    Pith/arXiv arXiv 2019

  15. [15]

    SDMetrics : Metrics for evaluating synthetic tabular data, 2023

    DataCebo, Inc. SDMetrics : Metrics for evaluating synthetic tabular data, 2023. Open-source Python library, Synthetic Data Vault (SDV) project, MIT Data to AI Lab

    2023

  16. [16]

    Fairness through awareness

    Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. Fairness through awareness. In Proceedings of the 3rd Innovations in Theoretical Computer Science Conference (ITCS), pages 214--226, 2012

    2012

  17. [17]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bj \"o rn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. arXiv:2012.09841

    Pith/arXiv arXiv 2021

  18. [18]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), volume 27, 2014

    2014

  19. [19]

    Borgwardt, Malte J

    Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Sch \"o lkopf, and Alexander Smola. A kernel two-sample test. Journal of Machine Learning Research, 13 0 (1): 0 723--773, 2012

    2012

  20. [20]

    Equality of opportunity in supervised learning

    Moritz Hardt, Eric Price, and Nathan Srebro. Equality of opportunity in supervised learning. In Advances in Neural Information Processing Systems (NeurIPS), volume 29, 2016. arXiv:1610.02413

    Pith/arXiv arXiv 2016

  21. [21]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022

    Pith/arXiv arXiv 2022

  22. [22]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS), volume 33, 2020. arXiv:2006.11239

    Pith/arXiv arXiv 2020

  23. [23]

    PATE-GAN : Generating synthetic data with differential privacy guarantees

    James Jordon, Jinsung Yoon, and Mihaela van der Schaar. PATE-GAN : Generating synthetic data with differential privacy guarantees. In Proceedings of the 7th International Conference on Learning Representations (ICLR), 2019

    2019

  24. [24]

    Cohen, and Adrian Weller

    James Jordon, Lukasz Szpruch, Florimond Houssiau, Mirko Bottarelli, Giovanni Cherubin, Carsten Maple, Samuel N. Cohen, and Adrian Weller. Synthetic data -- what, why and how? arXiv preprint arXiv:2205.03257, 2022

    Pith/arXiv arXiv 2022

  25. [25]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems (NeurIPS), volume 35, 2022. arXiv:2206.00364

    Pith/arXiv arXiv 2022

  26. [26]

    Analyzing and improving the training dynamics of diffusion models

    Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. arXiv:2312.02696. Introduces EDM2

    Pith/arXiv arXiv 2024

  27. [27]

    STaSy : Score-based tabular data synthesis

    Jayoung Kim, Chaejeong Lee, and Noseong Park. STaSy : Score-based tabular data synthesis. In Proceedings of the 11th International Conference on Learning Representations (ICLR), 2023. arXiv:2210.04018

    Pith/arXiv arXiv 2023

  28. [28]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In Proceedings of the 2nd International Conference on Learning Representations (ICLR), 2014. arXiv:1312.6114

    Pith/arXiv arXiv 2014

  29. [29]

    TabDDPM : Modelling tabular data with diffusion models

    Akim Kotelnikov, Dmitry Baranchuk, Ivan Rubachev, and Artem Babenko. TabDDPM : Modelling tabular data with diffusion models. In Proceedings of the 40th International Conference on Machine Learning (ICML), 2023. arXiv:2209.15421

    Pith/arXiv arXiv 2023

  30. [30]

    Kusner, Joshua Loftus, Chris Russell, and Ricardo Silva

    Matt J. Kusner, Joshua Loftus, Chris Russell, and Ricardo Silva. Counterfactual fairness. In Advances in Neural Information Processing Systems (NeurIPS), volume 30, 2017. arXiv:1703.06856

    Pith/arXiv arXiv 2017

  31. [31]

    Improved precision and recall metric for assessing generative models

    Tuomas Kynk \"a \"a nniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In Advances in Neural Information Processing Systems (NeurIPS), volume 32, 2019. arXiv:1904.06991

    Pith/arXiv arXiv 2019

  32. [32]

    A survey on datasets for fairness-aware machine learning

    Tai Le Quy, Arjun Roy, Vasileios Iosifidis, Wenbin Zhang, and Eirini Ntoutsi. A survey on datasets for fairness-aware machine learning. WIREs Data Mining and Knowledge Discovery, 12 0 (3), 2022. arXiv:2110.00530

    Pith/arXiv arXiv 2022

  33. [33]

    CoDi : Co-evolving contrastive diffusion models for mixed-type tabular synthesis

    Chaejeong Lee, Jayoung Kim, and Noseong Park. CoDi : Co-evolving contrastive diffusion models for mixed-type tabular synthesis. In Proceedings of the 40th International Conference on Machine Learning (ICML), 2023. arXiv:2304.12654

    Pith/arXiv arXiv 2023

  34. [34]

    GOGGLE : Generative modelling for tabular data by learning relational structure

    Tennison Liu, Zhaozhi Qian, Jeroen Berrevoets, and Mihaela van der Schaar. GOGGLE : Generative modelling for tabular data by learning relational structure. In Proceedings of the 11th International Conference on Learning Representations (ICLR), 2023

    2023

  35. [35]

    The variational fair autoencoder

    Christos Louizos, Kevin Swersky, Yujia Li, Max Welling, and Richard Zemel. The variational fair autoencoder. In Proceedings of the 4th International Conference on Learning Representations (ICLR), 2016. arXiv:1511.00830

    Pith/arXiv arXiv 2016

  36. [36]

    Learning adversarially fair and transferable representations

    David Madras, Elliot Creager, Toniann Pitassi, and Richard Zemel. Learning adversarially fair and transferable representations. In Proceedings of the 35th International Conference on Machine Learning (ICML), 2018. arXiv:1802.06309

    Pith/arXiv arXiv 2018

  37. [37]

    TabFairGDT : A fast fair tabular data generator using autoregressive decision trees

    Emmanouil Panagiotou, Beno \^ t Ronval, Arjun Roy, Ludwig Bothmann, Bernd Bischl, Siegfried Nijssen, and Eirini Ntoutsi. TabFairGDT : A fast fair tabular data generator using autoregressive decision trees. In Proceedings of the IEEE International Conference on Data Mining (ICDM), 2025. arXiv:2509.19927

    arXiv 2025

  38. [38]

    DiffuseVAE : Efficient, controllable and high-fidelity generation from low-dimensional latents

    Kushagra Pandey, Avideep Mukherjee, Piyush Rai, and Abhishek Kumar. DiffuseVAE : Efficient, controllable and high-fidelity generation from low-dimensional latents. Transactions on Machine Learning Research (TMLR), 2022. arXiv:2201.00308

    Pith/arXiv arXiv 2022

  39. [39]

    Causal fairness analysis: A causal toolkit for fair machine learning

    Drago Ple c ko and Elias Bareinboim. Causal fairness analysis: A causal toolkit for fair machine learning. Foundations and Trends in Machine Learning, 17 0 (3): 0 304--589, 2024. arXiv:2207.11385

    Pith/arXiv arXiv 2024

  40. [40]

    TabFairGAN : Fair tabular data generation with generative adversarial networks

    Amirarsalan Rajabi and Ozlem Ozmen Garibay. TabFairGAN : Fair tabular data generation with generative adversarial networks. Machine Learning and Knowledge Extraction, 4 0 (2): 0 488--501, 2022. arXiv:2109.00666

    Pith/arXiv arXiv 2022

  41. [41]

    Stochastic backpropagation and approximate inference in deep generative models

    Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In Proceedings of the 31st International Conference on Machine Learning (ICML), 2014. arXiv:1401.4082

    Pith/arXiv arXiv 2014

  42. [42]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. arXiv:2112.10752

    Pith/arXiv arXiv 2022

  43. [43]

    Mehdi S. M. Sajjadi, Olivier Bachem, Mario Lucic, Olivier Bousquet, and Sylvain Gelly. Assessing generative models via precision and recall. In Advances in Neural Information Processing Systems (NeurIPS), volume 31, 2018. arXiv:1806.00035

    Pith/arXiv arXiv 2018

  44. [44]

    Marshall, Severin Elvatun, Helga M nsted Br dersen, Sharmini Alagaratnam, Bjorn Edwin, and Narasimha Raghavan Veeraragavan

    Vegard Antun Vallevik, Aleksandar Babic, Serena E. Marshall, Severin Elvatun, Helga M nsted Br dersen, Sharmini Alagaratnam, Bjorn Edwin, and Narasimha Raghavan Veeraragavan. Can i trust my fake data -- a comprehensive quality assessment framework for synthetic tabular data in healthcare. International Journal of Medical Informatics, 185: 0 105413, 2024. ...

    Pith/arXiv arXiv 2024

  45. [45]

    DECAF : Generating fair synthetic data using causally-aware generative networks

    Boris van Breugel, Trent Kyono, Jeroen Berrevoets, and Mihaela van der Schaar. DECAF : Generating fair synthetic data using causally-aware generative networks. In Advances in Neural Information Processing Systems (NeurIPS), volume 34, 2021. arXiv:2110.12884

    Pith/arXiv arXiv 2021

  46. [46]

    Neural discrete representation learning

    Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems (NeurIPS), volume 30, 2017. arXiv:1711.00937

    Pith/arXiv arXiv 2017

  47. [47]

    Fairness definitions explained

    Sahil Verma and Julia Rubin. Fairness definitions explained. In Proceedings of the International Workshop on Software Fairness (FairWare), pages 1--7, 2018

    2018

  48. [48]

    CuTS : Customizable tabular synthetic data generation

    Mark Vero, Mislav Balunovi \'c , and Martin Vechev. CuTS : Customizable tabular synthetic data generation. In Proceedings of the 41st International Conference on Machine Learning (ICML), volume 235 of PMLR, 2024. arXiv:2307.03577

    Pith/arXiv arXiv 2024

  49. [49]

    In-processing modeling techniques for machine learning fairness: A survey

    Mingyang Wan, Daochen Zha, Ninghao Liu, and Na Zou. In-processing modeling techniques for machine learning fairness: A survey. ACM Transactions on Knowledge Discovery from Data, 17 0 (3), 2023. arXiv:2111.03015

    Pith/arXiv arXiv 2023

  50. [50]

    Fairlearn: Assessing and improving fairness of AI systems

    Hilde Weerts, Miroslav Dud \' k, Richard Edgar, Adrin Jalali, Roman Lutz, and Michael Madaio. Fairlearn: Assessing and improving fairness of AI systems. Journal of Machine Learning Research, 24 0 (257): 0 1--8, 2023. arXiv:2303.16626

    Pith/arXiv arXiv 2023

  51. [51]

    Fairness feedback loops: Training on synthetic data amplifies bias

    Sierra Wyllie, Ilia Shumailov, and Nicolas Papernot. Fairness feedback loops: Training on synthetic data amplifies bias. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT), 2024. arXiv:2403.07857

    Pith/arXiv arXiv 2024

  52. [52]

    FairGAN : Fairness-aware generative adversarial networks

    Depeng Xu, Shuhan Yuan, Lu Zhang, and Xintao Wu. FairGAN : Fairness-aware generative adversarial networks. In Proceedings of the 2018 IEEE International Conference on Big Data (Big Data), 2018. arXiv:1805.11202

    Pith/arXiv arXiv 2018

  53. [53]

    Modeling tabular data using conditional GAN

    Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Modeling tabular data using conditional GAN . In Advances in Neural Information Processing Systems (NeurIPS), volume 32, 2019. arXiv:1907.00503. Introduces CTGAN and TVAE

    Pith/arXiv arXiv 2019

  54. [54]

    Balanced mixed-type tabular data synthesis with diffusion models

    Zeyu Yang, Han Yu, Peikun Guo, Khadija Zanna, Xiaoxue Yang, and Akane Sano. Balanced mixed-type tabular data synthesis with diffusion models. Transactions on Machine Learning Research, 2025. arXiv:2404.08254

    Pith/arXiv arXiv 2025

  55. [55]

    Mixed-type tabular data synthesis with score-based diffusion in latent space

    Hengrui Zhang, Jiani Zhang, Balasubramaniam Srinivasan, Zhengyuan Shen, Xiao Qin, Christos Faloutsos, Huzefa Rangwala, and George Karypis. Mixed-type tabular data synthesis with score-based diffusion in latent space. In Proceedings of the 12th International Conference on Learning Representations (ICLR), 2024. arXiv:2310.09656

    Pith/arXiv arXiv 2024

  56. [56]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. arXiv:1801.03924

    Pith/arXiv arXiv 2018

  57. [57]

    Zilong Zhao, Aditya Kunar, Robert Birke, and Lydia Y. Chen. CTAB-GAN+ : Enhancing tabular data synthesis. Frontiers in Big Data, 6, 2023. arXiv:2204.00401

    Pith/arXiv arXiv 2023