Causal-Aware Tabular GANs with Reinforcement Learning

Tu Anh Hoang Nguyen; Dang Nguyen; Tri-Nhan Vo; Thuc Duy Le; Trung Le; Sunil Gupta

cs.LGOct 28, 2025

Causal-Aware Tabular GANs with Reinforcement Learning

Authors: Tu Anh Hoang Nguyen, Dang Nguyen, Tri-Nhan Vo, Thuc Duy Le, Trung Le, Sunil Gupta

Organizations: Applied Artificial Intelligence Initiative (A2I2), Deakin University, Australia · School of Computer Science and Information Technology, Adelaide University, Australia

Abstract

Existing tabular data generation methods primarily focus on matching statistical distributions between real and synthetic data, often overlooking the preservation of underlying causal relationships. As a result, generated samples may appear realistic while failing to maintain the causal structure required for reliable downstream analysis. We propose CA-GAN, a causal-aware generative framework for tabular data synthesis that explicitly incorporates causal knowledge into both the training and generation processes. CA-GAN first extracts a causal graph from real data to provide structural prior knowledge, then employs a graph-conditioned Conditional WGAN-GP whose sub-generators model variables according to their causal dependencies. More importantly, we introduce a reinforcement learning-based objective that treats causal graph discrepancy between real and synthetic data as a reward signal, enabling causal consistency to become an explicit optimization target during training rather than an implicit consequence of sampling order. Extensive experiments on 14 synthetic and real-world datasets demonstrate that CA-GAN consistently outperforms seven state-of-the-art baselines in causal preservation while achieving strong downstream utility, privacy preservation, and data quality. These results show that CA-GAN provides an effective and practical solution for generating high-quality synthetic tabular data that better respects underlying causal mechanisms.

Figures & tables

Figure 1: Our framework CA-GAN. First, we extract the causal graph

{\cal G}_{real}

from

{\cal D}_{real}

. Next, we train a WGAN-GP model with

\{G_{1},...,G_{M}\}

sub-generators and one discriminator

D

. Each

G_{j}

synthesizes variable

X_{j}

sequentially following the node ordering in

\mathcal{G}_{real}

. The input of

G_{j}

includes parent values

\mathbf{PA}_{j}

and noise

z_{j}

. We train

D

with a standard discriminator loss using fake samples

\hat{x}

and real samples

x

. Finally, we design a reward

R(\hat{x})

via the Structural Hamming Distance between

{\cal G}_{real}

and

{\cal G}_{fake}

extracted from

\hat{x}

. Using the Policy Gradient theorem, we construct the causal loss – the product of the log-probability of the generator output

\log p(\hat{x})

and

R(\hat{x})

. We train the generator with both adversarial loss and causal loss .

Panel of Figure 2 — Figure 2: True causal graphs (DAGs) used to construct 6 synthetic benchmark datasets.

Dataset	# of samples $N$	# of features $M$	Source
breast_cancer	569	30	scikit-learn
diabetes	768	8	OpenML
australian	690	14	OpenML
compas	4,010	10	UCI
phoneme	5,404	5	OpenML
bank	4,521	14	UCI

Table 1: Characteristics of 8 real-world benchmark datasets.

SHD ( $\downarrow$ )	Original	CTGAN	TVAE	TabDiff	Great	TapTap	Pred-LLM	Causal-TGAN	CA-GAN
4nodes_10k	2.00 (0.0)	3.67 (0.5)	3.33 (0.5)	4.67 (0.5)	3.00 (0.0)	3.00 (0.0)	3.00 (0.0)	2.00 (0.0)	2.00 (0.0)
5nodes_10k	0.00 (0.0)	5.67 (0.5)	5.67 (1.3)	7.00 (0.8)	7.33 (0.9)	2.00 (0.0)	6.00 (0.0)	4.00 (0.0)	4.00 (0.0)
6nodes_10k	5.00 (0.0)	8.67 (1.3)	8.33 (2.6)	8.33 (1.9)	9.67 (1.3)	11.00 (0.0)	5.00 (0.0)	6.00 (0.0)	5.67 (1.7)
4nodes_20k	2.00 (0.0)	3.00 (0.0)	2.67 (0.9)	4.00 (0.8)	4.00 (0.0)	5.00 (0.0)	3.00 (0.0)	2.00 (0.0)	2.00 (0.0)
5nodes_20k	4.00 (0.0)	5.00 (0.8)	4.67 (0.9)	6.00 (0.0)	5.67 (0.9)	6.00 (0.0)	6.00 (0.0)	4.00 (0.0)	4.00 (0.0)
6nodes_20k	4.00 (0.0)	8.00 (0.8)	7.67 (1.7)	8.67 (2.1)	8.33 (1.3)	8.00 (0.0)	9.00 (0.0)	8.00 (0.8)	5.00 (0.0)

Table 2: Causal preservation performance of generation methods. SHD

\pm

(std) are reported on 6 synthetic datasets. Bold and underline denote the best and second-best.

F1-score ( $\uparrow$ )	Original	CTGAN	TVAE	TabDiff	Great	TapTap	Pred-LLM	Causal-TGAN	CA-GAN
breast_ca	0.972 (0.0)	0.660 (0.2)	0.942 (0.0)	0.673 (0.0)	0.710 (0.1)	0.958 (0.0)	0.954 (0.0)	0.927 (0.1)	0.961 (0.0)
diabetes	0.604 (0.0)	0.375 (0.1)	0.633 (0.0)	0.181 (0.2)	0.360 (0.1)	0.621 (0.0)	0.562 (0.0)	0.618 (0.1)	0.640 (0.1)
australian	0.858 (0.0)	0.353 (0.2)	0.757 (0.0)	0.424 (0.1)	0.693 (0.0)	0.859 (0.0)	0.834 (0.0)	0.823 (0.1)	0.845 (0.0)
compas	0.278 (0.0)	0.116 (0.0)	0.210 (0.1)	0.028 (0.0)	0.094 (0.0)	0.323 (0.0)	0.257 (0.0)	0.250 (0.0)	0.260 (0.0)
phoneme	0.824 (0.0)	0.444 (0.0)	0.642 (0.0)	0.207 (0.1)	0.556 (0.1)	0.687 (0.0)	0.751 (0.0)	0.765 (0.0)	0.772 (0.0)
bank	0.400 (0.0)	0.075 (0.0)	0.319 (0.1)	0.006 (0.0)	0.061 (0.0)	0.289 (0.1)	0.307 (0.1)	0.347 (0.0)	0.357 (0.0)

Table 3: F1

\pm

(std) on 8 real-world datasets to evaluate utility preservation performance.

Figure 3: DCR distributions on the dataset bank . A very small DCR suggests the method may copy certain feature values from the original data (i.e., potential information leakage), whereas a very high DCR indicates unrealistic generated samples.

Figure 4: Average re-identification risk and quality score over 8 real-world datasets.

Figure 5: Influence of

\lambda

on CA-GAN’s performance over 8 real-world datasets.

Figure 6: The impact of different causal discovery methods. We report the average SHD (

\downarrow

) over 6 synthetic datasets. The consistent ranking across PC, GES, FCI, and BOSS demonstrates that CA-GAN’s benefit is not tied to any particular CD method.

DAG errors (%)	$\mathcal{G}_{real}$	Causal-TGAN	CA-GAN
0 (True mechanism)	$A\rightarrow S\leftarrow E$	2.00	2.00
20 (Missed parent)	$A\rightarrow S\hskip 17.07182ptE$	2.00	2.00
40 (No edges found)	$A\hskip 17.07182ptS\hskip 17.07182ptE$	2.33	2.00
60 (Wrong chain)	$A\rightarrow E\rightarrow S$	2.67	2.33
80 (Effect as cause)	$A\leftarrow S\hskip 17.07182ptE$	3.00	3.00
100 (Total mismatch)	$A\leftarrow S\rightarrow E$	3.00	3.00

Table 4: DAG-error and SHD (

\downarrow

) of Causal-TGAN and CA-GAN.

Figure 7: Average runtime ( in minutes ) of baselines across eight real-world datasets.

Explore similar work

Improving TabPFN's Synthetic Data Generation by Integrating Causal Structure
Mar 10, 2026Davide Tugnoli, Andrea De Lorenzo, Marco Virgolin +1Synthetic Tabular Data Tabular Prior-Data Fitted Network
Department of Mathematics, Informatics and Geosciences, University of Trieste, Trieste, Italy · InSilicoTrials Technologies, Riva Grumula 2, 34123 Trieste, Italy · Department of Engineering and Architecture, University of Trieste, Trieste, Italy +3
Generative Synthetic Data for Causal Inference: Pitfalls, Remedies, and Opportunities
Apr 26, 2026Yichen XuCausal Inferences Synthetic Data
University of California, Berkeley
Causality for Tabular Data Synthesis: A High-Order Structure Causal Benchmark Framework
Jun 12, 2024Zineb Senane, Axel Karlsson, Lele Cao +6Synthetic Tabular Data Causal Modeling
KTH Royal Institute of Technology · Stockholm, Sweden · King AI Labs, Microsoft +2

Mar 10, 2026, cs.LGJ/K move · Enter open · S save

Improving TabPFN's Synthetic Data Generation by Integrating Causal Structure

Davide Tugnoli, Andrea De Lorenzo, Marco Virgolin, Giovanni Cinà

Department of Mathematics, Informatics and Geosciences, University of Trieste, Trieste, Italy · InSilicoTrials Technologies, Riva Grumula 2, 34123 Trieste, Italy · Department of Engineering and Architecture, University of Trieste, Trieste, Italy+3

Synthetic Tabular Data Tabular Prior-Data Fitted Network Synthetic Data Directed Acyclic Graph

Open paper AskPDF arXiv

Synthetic tabular data generation addresses data scarcity and privacy constraints in a variety of domains. Tabular Prior-Data Fitted Network (TabPFN), a recent foundation model for tabular data, has been shown capable of generating high-quality synthetic tabular data. However, TabPFN is autoregressive: features are generated sequentially by conditioning on the previous ones, depending on the order in which they appear in the input data. We demonstrate that when the feature order conflicts with causal structure, the model produces spurious correlations that impair its ability to generate synthetic data and preserve causal effects. We address this limitation by integrating causal structure into TabPFN's generation process through two complementary approaches: Directed Acyclic Graph (DAG)-aware conditioning, which samples each variable given its causal parents, and a partially directed acyclic graph (PDAG)-based strategy for scenarios with partial causal knowledge. We evaluate these approaches on controlled benchmarks and six CSuite datasets, assessing structural fidelity, distributional quality, and Average Treatment Effect (ATE) preservation. Across most settings, DAG-aware conditioning improves the quality and stability of synthetic data relative to vanilla TabPFN. Under partial causal knowledge, the oracle partially directed acyclic graph (oracle-PDAG), which orients only the edges into the colliders, shows moderate gains, while the benefit of a Completed Partially Directed Acyclic Graph (CPDAG) discovered from data depends on how well the causal structure is recovered. These results indicate that reliable causal structure, even partial, can be injected into TabPFN at inference time, without parameter updates, to improve synthetic data quality.

Abstract

Explore similar work

Improving TabPFN's Synthetic Data Generation by Integrating Causal Structure

Generative Synthetic Data for Causal Inference: Pitfalls, Remedies, and Opportunities

Causality for Tabular Data Synthesis: A High-Order Structure Causal Benchmark Framework