arxiv: 2604.22337 · v1 · submitted 2026-04-24 · 💻 cs.LG

Recognition: unknown

TabSCM: A practical Framework for Generating Realistic Tabular Data

Sven Jacob , Bardh Prenkaj , Weijia Shao , Gjergji Kasneci

Authors on Pith no claims yet

Pith reviewed 2026-05-08 12:30 UTC · model grok-4.3

classification 💻 cs.LG

keywords tabular data generationcausal structurediffusion modelsgradient boosted treescounterfactual generationstructural causal modelsmixed-type datadata privacy

0 comments

The pith

TabSCM generates tabular data by learning conditional distributions along a discovered causal graph rather than matching statistics alone.

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

Tabular data generators often reproduce marginal statistics but overlook causal dependencies, so downstream models pick up spurious correlations. TabSCM begins with a causal graph discovered by any algorithm, orients it into a DAG, fits root marginals, and then learns the remaining conditional distributions using diffusion models for continuous variables and gradient-boosted trees for categorical ones. Ancestral sampling from this structure produces records that respect the causal order, supports exact counterfactuals, and reduces invalid outputs. On seven public datasets the approach matches or exceeds current GAN, diffusion, and LLM generators in fidelity, downstream task performance, and privacy metrics while running substantially faster.

Core claim

TabSCM orients a CPDAG into a DAG, estimates root-node marginals with kernel density estimation or frequency counts, and fits topologically ordered structural assignments by training conditional diffusion models on continuous child nodes and gradient-boosted trees on categorical child nodes; ancestral sampling then yields causally coherent records and enables precise interventional queries.

What carries the argument

A CPDAG-derived DAG that decomposes the joint distribution into explicit conditional structural assignments learned by mixed-type models.

If this is right

Generated data exhibits lower rates of rule violations than non-causal baselines.
Downstream models trained on the data achieve higher utility than those trained on outputs from GAN, diffusion, or LLM generators.
Privacy risk metrics remain comparable or better while causal interventions stay robust.
Generation runs up to 583 times faster than pure diffusion models because sampling decomposes into independent conditional steps.
The explicit equations expose interpretable parameters for auditing fairness and simulating policy changes.

Where Pith is reading between the lines

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

Substituting domain-expert graphs for the off-the-shelf discovery step would let TabSCM serve as a policy simulator in regulated fields such as healthcare or finance.
The same decomposition could be applied to other structured generative tasks, for example time-series or graph data, by replacing the tabular conditionals with appropriate sequence or graph learners.
Sensitivity tests that vary the accuracy of the input CPDAG would quantify how much TabSCM's gains depend on causal discovery quality.

Load-bearing premise

The causal graph recovered by standard discovery algorithms is close enough to the true structure that the fitted conditionals remain valid under intervention.

What would settle it

Generate data with TabSCM on a synthetic dataset whose ground-truth DAG is known, then repeat the experiment after feeding the discovery algorithm a deliberately corrupted version of that DAG and check whether the fidelity, rule-violation, and intervention advantages disappear.

Figures

Figures reproduced from arXiv: 2604.22337 by Bardh Prenkaj, Gjergji Kasneci, Sven Jacob, Weijia Shao.

**Figure 1.** Figure 1: Shows average error (Eq. 3) and average training time of TabSCM against view at source ↗

**Figure 2.** Figure 2: Minimal example of a system of four observed variables Xi , and corresponding exogenous variables ϵi for i = 1, 2, 3, 4. The causal relationships and interactions of the observed variables are illustrated on the left-hand side (causal graph G). On the right-hand side, we describe the SCM for the associated causal graph G on the left. where V = {X1, X2, . . . , Xd} is the set of observed variables (nodes)… view at source ↗

**Figure 3.** Figure 3: The conceptual framework of our proposed method, including i) causal view at source ↗

**Figure 4.** Figure 4: Shows the mean density error, correlation error, and AUC scores for view at source ↗

**Figure 5.** Figure 5: Illustrates how the number of epochs influences density estimation error view at source ↗

**Figure 6.** Figure 6: Shows False Negative Ratio (FNR) and False Positive Ratio (FPR) for view at source ↗

**Figure 7.** Figure 7: Shows the SHAP values of the same model trained on real data and view at source ↗

**Figure 8.** Figure 8: Shows the mean absolute SHAP values for the fitted structural assignment view at source ↗

**Figure 9.** Figure 9: Shows the marginal distribution of the free attributes for counterfactual view at source ↗

read the original abstract

Most tabular-data generators match marginal statistics yet ignore causal structure, leading downstream models to learn spurious or unfair patterns. We present TabSCM, a mixed-type generator that preserves those causal dependencies. Starting from a Completed Partially Directed Acyclic Graph (CPDAG) found by any causal structure discovery algorithm, TabSCM (i) orients edges to a DAG, (ii) fits root-node marginals with KDE or categorical frequencies, and (iii) learns topologically ordered structural assignments. Such assignments are achieved using conditional diffusion models for continuous variables as child nodes and gradient-boosted trees for categorical ones. Ancestral sampling yields semantically valid records and enables exact counterfactual queries. On seven public datasets, encompassing healthcare, finance, housing, environment, TabSCM matches or surpasses state-of-the-art GAN, diffusion, and LLM baselines in statistical fidelity, downstream utility, and privacy risk, while also cutting rule-violation rates and providing causally meaningful and robust conditional interventions. Because generation is decomposed into explicit equations, it runs up to 583$\times$ faster than diffusion-only models and exposes interpretable knobs for fairness auditing and policy simulation, making TabSCM a practical choice for realism, explainability, and causal soundness.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

TabSCM mixes causal discovery with type-specific conditionals and ancestral sampling for tabular generation, but the abstract's superiority claims lack numbers and the causal validity rests on untested graph accuracy.

read the letter

TabSCM starts from a CPDAG produced by any off-the-shelf causal discovery tool, orients it to a DAG, models root marginals with KDE or frequencies, then learns structural equations with conditional diffusion for continuous children and gradient-boosted trees for categorical ones. Ancestral sampling produces rows that respect the topology and supports direct interventions without retraining. The concrete pipeline that combines these pieces for mixed-type data and counterfactual queries is not laid out in the cited prior work. The decomposition into explicit steps also gives speed gains up to hundreds of times faster than full diffusion models and makes the generation process more interpretable for auditing. Those are the practical upsides if the results hold. The abstract asserts that the method matches or beats GAN, diffusion, and LLM baselines on seven public datasets for fidelity, utility, privacy, and rule compliance while adding causal interventions. No tables, error bars, or ablation numbers appear in the provided description, so the strength of those claims cannot be judged from what is shown. The larger gap is that everything downstream of the graph depends on the discovered CPDAG being close enough to the true structure. The write-up states the method works with any discovery algorithm but reports no sensitivity checks, no comparisons across discovery methods, and no controlled edge-error injections to see how fidelity or intervention quality degrades. That assumption is central to the causal claims and remains unexamined. This work is aimed at people who generate synthetic tabular data for high-stakes domains where preserving causal dependencies matters more than pure statistical matching. A reader focused on causal generative models or practical synthetic data pipelines would find the pipeline description and speed angle useful. It deserves a serious referee because the approach is concrete and implementable, even though the current evidence base needs the missing quantitative results and robustness tests before the causal advantages can be taken as established. I would send it to review with requests for the tables and the sensitivity analysis on discovery errors.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces TabSCM, a mixed-type tabular data generator that incorporates causal structure. It takes a CPDAG from any off-the-shelf causal discovery algorithm, orients it to a DAG, models root marginals via KDE or frequencies, and learns topologically ordered structural assignments (conditional diffusion for continuous children, gradient-boosted trees for categorical). Ancestral sampling produces records; the approach is claimed to match or exceed GAN/diffusion/LLM baselines on seven public datasets in statistical fidelity, downstream utility, privacy, rule-violation rates, and causal intervention quality, while running up to 583× faster due to its explicit decomposition.

Significance. If the reported empirical gains hold under rigorous verification and the causal assumptions prove robust, TabSCM would provide a practical, interpretable alternative to black-box generators for domains such as healthcare and finance. The explicit structural decomposition enabling fast sampling, counterfactual queries, and fairness auditing is a clear strength over purely implicit models.

major comments (3)

[§4 and §5] §4 (Experimental Setup) and §5 (Results): the abstract and introduction assert superiority on seven datasets in fidelity, utility, and rule-violation metrics, yet no tables, error bars, statistical significance tests, or ablation studies isolating the causal component versus non-causal baselines are referenced; without these, the central empirical claim cannot be evaluated.
[§3.1] §3.1 (Causal Structure Learning): the validity of learned conditionals and 'causally meaningful interventions' rests on the supplied CPDAG being close to the true structure, but the manuscript reports no sensitivity analysis, no comparison across multiple discovery algorithms, and no controlled edge-error injection experiments to quantify degradation in fidelity or intervention quality.
[§3.3] §3.3 (Structural Assignments): the claim that diffusion and GBT assignments remain valid under ancestral sampling and do-calculus interventions is load-bearing for the causal soundness argument, yet no formal justification or empirical check against ground-truth DAGs is supplied when the discovery step errs.

minor comments (2)

[§3.2] Notation for the oriented DAG after CPDAG processing is introduced without an explicit equation or pseudocode step; a small diagram or numbered procedure would improve clarity.
[§5] The 583× speedup claim is stated without a corresponding table listing wall-clock times against each baseline on identical hardware.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing TabSCM's potential as a practical, interpretable alternative to black-box generators. We address each major comment below, committing to revisions that strengthen the empirical claims and robustness analysis while preserving the manuscript's core contributions.

read point-by-point responses

Referee: [§4 and §5] §4 (Experimental Setup) and §5 (Results): the abstract and introduction assert superiority on seven datasets in fidelity, utility, and rule-violation metrics, yet no tables, error bars, statistical significance tests, or ablation studies isolating the causal component versus non-causal baselines are referenced; without these, the central empirical claim cannot be evaluated.

Authors: We agree that the presentation of results can be strengthened for clarity and rigor. The current manuscript includes comparative evaluations across the seven datasets in §5, but we acknowledge the absence of consolidated tables with means and standard deviations, error bars on figures, formal statistical significance tests, and explicit ablations isolating the causal orientation step. In the revision we will add: (i) a summary table reporting all metrics with standard deviations over 5 random seeds and paired t-test p-values against baselines; (ii) error bars on all fidelity and utility plots; and (iii) a dedicated ablation subsection comparing TabSCM against a non-causal variant that uses the same diffusion/GBT components but ignores the learned DAG topology. These additions will make the central claims directly verifiable. revision: yes
Referee: [§3.1] §3.1 (Causal Structure Learning): the validity of learned conditionals and 'causally meaningful interventions' rests on the supplied CPDAG being close to the true structure, but the manuscript reports no sensitivity analysis, no comparison across multiple discovery algorithms, and no controlled edge-error injection experiments to quantify degradation in fidelity or intervention quality.

Authors: This concern is well-founded. The manuscript treats the CPDAG as an input from any off-the-shelf algorithm and focuses on the subsequent generation pipeline. To quantify robustness we will add a new subsection in §5 that (a) runs TabSCM with CPDAGs produced by both the PC algorithm and NOTEARS on the same datasets, (b) reports fidelity, utility, and intervention metrics for each, and (c) includes a controlled edge-error injection study on synthetic data with known ground-truth DAGs, measuring degradation as a function of flipped or missing edges. These experiments will be presented with the same metrics used in the main results. revision: yes
Referee: [§3.3] §3.3 (Structural Assignments): the claim that diffusion and GBT assignments remain valid under ancestral sampling and do-calculus interventions is load-bearing for the causal soundness argument, yet no formal justification or empirical check against ground-truth DAGs is supplied when the discovery step errs.

Authors: We partially agree. When the oriented DAG is correct, ancestral sampling is valid by the topological ordering and the explicit structural equations; do-interventions are realized by clamping the intervened node and resampling its descendants, which follows standard causal semantics. However, the manuscript does not supply a formal proof of validity under discovery errors nor ground-truth checks on erroneous CPDAGs. In revision we will (i) add a short discussion in §3.3 clarifying the assumption that the supplied CPDAG is sufficiently accurate and (ii) include an empirical study on synthetic ground-truth DAGs that injects controlled errors and reports the resulting drop in fidelity and intervention quality. This provides the requested empirical check without overstating robustness. revision: partial

Circularity Check

0 steps flagged

No circularity: procedural generative method with external structure input and empirical evaluation

full rationale

TabSCM is defined as a pipeline that ingests an externally supplied CPDAG (from any off-the-shelf discovery algorithm), orients it to a DAG, fits root marginals via KDE/frequencies, and learns conditional structural assignments (conditional diffusion for continuous, GBTs for categorical) in topological order. Generation proceeds by ancestral sampling. All performance claims (fidelity, utility, privacy, rule violations, interventions) are obtained by running the fitted generator on held-out public datasets and comparing to external baselines. No equation or claim reduces a derived quantity to a fitted parameter by construction, no uniqueness theorem is imported via self-citation, and no ansatz is smuggled; the causal soundness claim rests on the (explicitly stated) assumption that the input CPDAG is sufficiently accurate, which is an external modeling choice rather than an internal tautology.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The framework rests on standard causal assumptions plus several modeling choices whose impact is not quantified in the abstract.

free parameters (2)

KDE bandwidth for root marginals
Chosen to fit continuous root-node distributions; value not stated.
Diffusion noise schedule and tree hyperparameters
Learned or tuned per conditional model; no count or selection procedure given.

axioms (2)

domain assumption The input CPDAG from any causal discovery algorithm is sufficiently accurate to support valid conditional distributions under intervention
Invoked when the method orients edges and learns parent-conditioned models.
standard math Topological ordering permits sequential ancestral sampling without cycles
Relies on the DAG property after orientation.

pith-pipeline@v0.9.0 · 5520 in / 1543 out tokens · 32145 ms · 2026-05-08T12:30:51.510903+00:00 · methodology

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Tabular Foundation Model for Generative Modelling
cs.LG 2026-05 unverdicted novelty 5.0

TabFORGE generates high-quality synthetic tabular data by leveraging pretrained causality-aware representations in a two-stage diffusion-decoder architecture that mitigates latent distribution shifts.

Reference graph

Works this paper leans on

40 extracted references · 5 canonical work pages · cited by 1 Pith paper · 1 internal anchor

[1]

In: International conference on machine learning

Alaa, A., Van Breugel, B., Saveliev, E.S., Van Der Schaar, M.: How faithful is your synthetic data? sample-level metrics for evaluating and auditing generative models. In: International conference on machine learning. pp. 290–306. PMLR (2022)

2022
[2]

Applied Sciences12(9), 4619 (2022)

Barbierato, E., Vedova, M.L.D., Tessera, D., Toti, D., Vanoli, N.: A method- ology for controlling bias and fairness in synthetic data generation. Applied Sciences12(9), 4619 (2022)

2022
[3]

Becker and R

Becker, B., Kohavi, R.: Adult. UCI Machine Learning Repository (1996), DOI: https://doi.org/10.24432/C5XW20

work page doi:10.24432/c5xw20 1996
[4]

UCI Machine Learning Repository (2004), DOI: https://doi.org/10.24432/C52C8B

Bock, R.: MAGIC Gamma Telescope. UCI Machine Learning Repository (2004), DOI: https://doi.org/10.24432/C52C8B

work page doi:10.24432/c52c8b 2004
[5]

In: The Eleventh International Conference on Learning Representations (2023),https://openreview

Borisov,V.,Sessler,K.,Leemann,T.,Pawelczyk,M.,Kasneci,G.:Language models are realistic tabular data generators. In: The Eleventh International Conference on Learning Representations (2023),https://openreview. net/forum?id=cEygmQNOeI

2023
[6]

Journal of artificial intelligence research16, 321–357 (2002)

Chawla, N.V., Bowyer, K.W., Hall, L.O., Kegelmeyer, W.P.: Smote: syn- thetic minority over-sampling technique. Journal of artificial intelligence research16, 321–357 (2002)

2002
[7]

Nature Biomedical Engineering5(6), 493–497 (2021)

Chen, R.J., Lu, M.Y., Chen, T.Y., Williamson, D.F., Mahmood, F.: Synthetic data in machine learning for medicine and healthcare. Nature Biomedical Engineering5(6), 493–497 (2021)

2021
[8]

In: Pro- ceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining

Chen, T., Guestrin, C.: Xgboost: A scalable tree boosting system. In: Pro- ceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining. pp. 785–794 (2016)

2016
[9]

Chickering,D.M.:Optimalstructureidentificationwithgreedysearch.Jour- nal of machine learning research3(Nov), 507–554 (2002)

2002
[10]

Advances in neural information processing systems34, 8780–8794 (2021)

Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems34, 8780–8794 (2021)

2021
[11]

Esteban, C., Hyland, S.L., Rätsch, G.: Real-valued (medical) time series generationwithrecurrentconditionalgans.arXivpreprintarXiv:1706.02633 (2017)

work page arXiv 2017
[12]

Neurocomputing321, 321–331 (2018)

Frid-Adar, M., Diamant, I., Klang, E., Amitai, M., Goldberger, J., Greenspan, H.: Gan-based synthetic medical image augmentation for in- creased cnn performance in liver lesion classification. Neurocomputing321, 321–331 (2018)

2018
[13]

Annals of statistics pp

Friedman, J.H.: Greedy function approximation: a gradient boosting ma- chine. Annals of statistics pp. 1189–1232 (2001)

2001
[14]

BMC medical research methodology20, 1–40 (2020)

Goncalves, A., Ray, P., Soper, B., Stevens, J., Coyle, L., Sales, A.P.: Gen- eration and evaluation of synthetic patient data. BMC medical research methodology20, 1–40 (2020)

2020
[15]

In: 2008 IEEE international joint con- 24 S

He, H., Bai, Y., Garcia, E.A., Li, S.: Adasyn: Adaptive synthetic sampling approach for imbalanced learning. In: 2008 IEEE international joint con- 24 S. Jacob et al. ference on neural networks (IEEE world congress on computational intelli- gence). pp. 1322–1328. Ieee (2008)

2008
[16]

Com- puter Vision and Machine Intelligence in Medical Image Analysis (2019), https://api.semanticscholar.org/CorpusID:202835967

Islam, M.M.F., Ferdousi, R., Rahman, S., Bushra, H.Y.: Likelihood pre- diction of diabetes at early stage using data mining techniques. Com- puter Vision and Machine Intelligence in Medical Image Analysis (2019), https://api.semanticscholar.org/CorpusID:202835967

2019
[17]

In: International conference on machine learning

Johansson, F., Shalit, U., Sontag, D.: Learning representations for coun- terfactual inference. In: International conference on machine learning. pp. 3020–3029. PMLR (2016)

2016
[18]

In: International conference on learning representations (2018)

Jordon, J., Yoon, J., Van Der Schaar, M.: Pate-gan: Generating synthetic data with differential privacy guarantees. In: International conference on learning representations (2018)

2018
[19]

In: International Conference on Learning Representations (2018),https:// openreview.net/forum?id=BJE-4xW0W

Kocaoglu, M., Snyder, C., Dimakis, A.G., Vishwanath, S.: CausalGAN: Learning causal implicit generative models with adversarial training. In: International Conference on Learning Representations (2018),https:// openreview.net/forum?id=BJE-4xW0W

2018
[20]

In: International Conference on Machine Learning

Kotelnikov, A., Baranchuk, D., Rubachev, I., Babenko, A.: Tabddpm: Mod- elling tabular data with diffusion models. In: International Conference on Machine Learning. pp. 17564–17579. PMLR (2023)

2023
[21]

5 pollution: severity, weather impact, apec and winter heating

Liang, X., Zou, T., Guo, B., Li, S., Zhang, H., Zhang, S., Huang, H., Chen, S.X.: Assessing beijing’s pm2. 5 pollution: severity, weather impact, apec and winter heating. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences471(2182), 20150257 (2015)

2015
[22]

In: The Eleventh International Conference on Learning Representations (2023)

Liu, T., Qian, Z., Berrevoets, J., van der Schaar, M.: Goggle: Generative modelling for tabular data by learning relational structure. In: The Eleventh International Conference on Learning Representations (2023)

2023
[23]

Advances in neural information processing systems30(2017)

Louizos, C., Shalit, U., Mooij, J.M., Sontag, D., Zemel, R., Welling, M.: Causal effect inference with deep latent-variable models. Advances in neural information processing systems30(2017)

2017
[24]

In: Proceedings of the 2020 conference on fairness, accountability, and transparency

Mothilal, R.K., Sharma, A., Tan, C.: Explaining machine learning classifiers through diverse counterfactual explanations. In: Proceedings of the 2020 conference on fairness, accountability, and transparency. pp. 607–617 (2020)

2020
[25]

Statistics & Proba- bility Letters33(3), 291–297 (1997)

Pace, R.K., Barry, R.: Sparse spatial autoregressions. Statistics & Proba- bility Letters33(3), 291–297 (1997)

1997
[26]

Park, N., Mohammadi, M., Gorde, K., Jajodia, S., Park, H., Kim, Y.: Data synthesis based on generative adversarial networks. Proc. VLDB Endow. 11(10), 1071–1083 (Jun 2018).https://doi.org/10.14778/3231751. 3231757,https://doi.org/10.14778/3231751.3231757

work page doi:10.14778/3231751 2018
[27]

Cambridge Univer- sity Press, New York (2000)

Pearl, J.: Causality: Models, Reasoning and Inference. Cambridge Univer- sity Press, New York (2000)

2000
[28]

In: Proceedings of the Twenty-Eighth Conference on Uncertainty in Artificial Intelligence

Pearl, J.: The do-calculus revisited. In: Proceedings of the Twenty-Eighth Conference on Uncertainty in Artificial Intelligence. p. 3–11. UAI’12, AUAI Press, Arlington, Virginia, USA (2012)

2012
[29]

The MIT Press (2017) TabSCM: A practical Framework for Generating Realistic Tabular Data 25

Peters, J., Janzing, D., Schölkopf, B.: Elements of causal inference: founda- tions and learning algorithms. The MIT Press (2017) TabSCM: A practical Framework for Generating Realistic Tabular Data 25

2017
[30]

Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018)

2018
[31]

Hierarchical Text-Conditional Image Generation with CLIP Latents

Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchi- cal text-conditional image generation with clip latents. arXiv preprint arXiv:2204.061251(2), 3 (2022)

work page internal anchor Pith review arXiv 2022
[32]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High- resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022)

2022
[33]

In: The Thirteenth International Conference on Learning Representations (2025), https://openreview.net/forum?id=swvURjrt8z

Shi, J., Xu, M., Hua, H., Zhang, H., Ermon, S., Leskovec, J.: Tabd- iff: a mixed-type diffusion model for tabular data generation. In: The Thirteenth International Conference on Learning Representations (2025), https://openreview.net/forum?id=swvURjrt8z

2025
[34]

MIT press (2000)

Spirtes, P., Glymour, C.N., Scheines, R.: Causation, prediction, and search. MIT press (2000)

2000
[35]

Big Data & Society4(2), 2053951717743530 (2017)

Veale, M., Binns, R.: Fairer machine learning in the real world: Mitigating discrimination without collecting sensitive data. Big Data & Society4(2), 2053951717743530 (2017)

2017
[36]

In: ICLR Work- shop on Deep Generative Models for Highly Structured Data (2022),https: //openreview.net/forum?id=BEhxCh4dvW5

Wen, B., Cao, Y., Yang, F., Subbalakshmi, K., Chandramouli, R.: Causal- TGAN: Modeling tabular data using causally-aware GAN. In: ICLR Work- shop on Deep Generative Models for Highly Structured Data (2022),https: //openreview.net/forum?id=BEhxCh4dvW5

2022
[37]

In: 2018 IEEE international conference on big data (big data)

Xu, D., Yuan, S., Zhang, L., Wu, X.: Fairgan: Fairness-aware generative adversarial networks. In: 2018 IEEE international conference on big data (big data). pp. 570–575. IEEE (2018)

2018
[38]

Advances in neural information process- ing systems32(2019)

Xu, L., Skoularidou, M., Cuesta-Infante, A., Veeramachaneni, K.: Modeling tabular data using conditional gan. Advances in neural information process- ing systems32(2019)

2019
[39]

In: The Twelfth International Conference on Learning Representations (2024),https://openreview.net/forum?id= 4Ay23yeuz0

Zhang, H., Zhang, J., Shen, Z., Srinivasan, B., Qin, X., Faloutsos, C., Rangwala, H., Karypis, G.: Mixed-type tabular data synthesis with score- based diffusion in latent space. In: The Twelfth International Conference on Learning Representations (2024),https://openreview.net/forum?id= 4Ay23yeuz0

2024
[40]

Advances in neural infor- mation processing systems31(2018)

Zheng, X., Aragam, B., Ravikumar, P.K., Xing, E.P.: Dags with no tears: Continuous optimization for structure learning. Advances in neural infor- mation processing systems31(2018)

2018