CTGAN for Tabular Data Synthesis

• CTGAN is a conditional generative adversarial network that synthesizes realistic tabular data by integrating both continuous and discrete features through targeted conditioning.
• It employs mode-specific normalization with Gaussian Mixture Models to accurately represent multi-modal distributions and ensure reversible data preprocessing.
• Advanced variants like CTG-KrEW optimize high-cardinality and sparse representations, reducing memory and computational overhead while preserving data fidelity.

A Conditional Tabular Generative Adversarial Network (CTGAN) is a class of generative models specifically designed for synthesizing realistic tabular data with mixed continuous and discrete features, especially in settings where distributions are multi-modal or categories are imbalanced. Since its introduction, CTGAN and its variants have become foundational for privacy-respecting data synthesis, rare class oversampling, causal inference, and structured semantic content generation, with extensive empirical validation across public benchmarks and real-world datasets (Xu et al., 2019, Samanta et al., 2024, Menssouri et al., 9 Feb 2025, Nasution et al., 25 Feb 2026, Grassi et al., 27 Mar 2026).

1. Core Architecture and Conditional Training

CTGAN extends the classical GAN paradigm by explicitly conditioning the generator on a chosen discrete column, enabling targeted synthesis of rare or underrepresented categories. Each training step involves selecting a discrete attribute and sampling its value according to its marginal frequency; the generator receives both a latent vector $z \sim \mathcal{N}(0, I)$ and this one-hot–encoded condition vector $c$. The main objective is the standard minimax game, optionally realized with a Wasserstein GAN plus gradient penalty to improve stability:

$$V(D,G)=\mathbb{E}_{x\sim P_\text{data}}[\log D(x|c)] + \mathbb{E}_{z\sim P_z}[\log(1-D(G(z|c)))]$$

$$\begin{aligned} L_D & = -\mathbb{E}_{x\sim P_\text{data}}[D(x)] + \mathbb{E}_{z\sim P_z}[D(G(z,c))] + \lambda\,\mathbb{E}_{\hat{x}\sim P_{\hat{x}}}\left[(\|\nabla_{\hat{x}}D(\hat{x})\|_2-1)^2\right] \ L_G & = -\mathbb{E}_{z\sim P_z}[D(G(z,c))] \end{aligned}$$

To enforce that the generated output conforms to the selected condition $c$, a cross-entropy classification term is often added to the generator loss for discrete columns:

$$L_\text{cond} = \mathbb{E}_{z, c}[\text{CE}(c, G_c(z, c))]$$

The generator is implemented as an MLP with two or more fully connected layers (plus BatchNorm and ReLU activations), outputting continuous values via a tanh or identity function, while discrete attributes are typically modeled with Gumbel-Softmax heads. The discriminator (or critic) is similarly realized as an MLP or convolutional architecture that processes concatenated real or synthetic samples along with the same condition vector, outputting a scalar authenticity estimate. PacGAN-style mini-batch discrimination, where the critic receives “packs” of $m$ rows, is employed to mitigate mode collapse (Xu et al., 2019, Samanta et al., 2024).

2. Mode-Specific Normalization and Mixed-Type Representation

A defining technical feature of CTGAN is its treatment of continuous variables exhibiting arbitrarily multi-modal distributions. For each continuous column, a Gaussian Mixture Model (GMM) is fit during preprocessing. Each value is assigned a mode via the posterior over mixture components, then standardized within that mode:

$\tilde{x} = \frac{x - \mu_{k^*}}{\sigma_{k^*}}$

where $k^* = \arg\max_k p_k(x)$, with $p_k(x)$ denoting the GMM component likelihoods. The preprocessed vector includes both one-hot mode indicators and normalized residuals, ensuring invertibility and accurate reproduction of the original distribution at inference (Xu et al., 2019, Menssouri et al., 9 Feb 2025, Nasution et al., 25 Feb 2026, Grassi et al., 27 Mar 2026).

Discrete columns are one-hot encoded, while binary columns are mapped to $c$0. The complete input/output row vector thus concatenates normalized continuous attributes, mode indicators, and one-hot discrete codes.

3. Enhanced Conditional Synthesis and Efficient Extensions

While CTGAN reliably synthesizes tabular data across a range of complexity, canonical implementations encounter memory and runtime bottlenecks under extremely high-cardinality or sparse multi-hot representations. The CTG-KrEW framework extends CTGAN to efficiently handle variable-length collections of semantically related words (as in freelancer “skills” columns) via the following innovation pipeline (Samanta et al., 2024):

• Word2Vec Embedding: A skip-gram model is trained on pseudo-sentences reflecting co-occurrence of skills, producing dense vector embeddings for each skill token.
• K-Means Clustering: The vocab embedding space is partitioned into $c$1 clusters (empirically chosen, e.g. $c$2). Each record’s skillset is encoded as a $c$3-dimensional “cluster-count” vector $c$4.
• Tabular Transformation: The original high-dimensional multi-hot skill columns are replaced by their cluster-count representations, concatenated to the remaining columns, yielding a compact input matrix of size $c$5.
• Standard CTGAN Training: The generator and discriminator operate on the transformed table, treating cluster-counts as continuous features.

Synthetic skillsets are decoded by rounding/flooring the generated cluster-counts and sampling words within each cluster by membership frequency, ensuring co-occurrence and semantic fidelity. CTG-KrEW achieves approximate entropy and contextual similarity parity with the much more resource-intensive multi-hot encoding approaches, while reducing memory by $c$6 and CPU consumption by $c$7 in experiments (Samanta et al., 2024).

4. Performance Metrics and Empirical Comparisons

CTGAN and its cluster-embedding derivatives are evaluated via multi-faceted measures:

• Skillset Variability: Shannon entropy over the distribution of synthesized attribute sets.
• Contextual Similarity: Cosine similarity between generated and real skillsets (using a BERT-based SentenceTransformer).
• Frequency Distribution: Kullback-Leibler divergence between real and synthetic marginal token frequencies.
• Associativity: Pearson correlation of skill co-occurrence matrices.
• Quality of Remaining Columns: Histogram or PDF overlays on categorical/continuous columns.

CTG-KrEW matches or exceeds baseline CTGAN-MHE on entropy and KL divergence, with Pearson correlation (co-occurrence) reaching $c$8–$c$9 versus $$V(D,G)=\mathbb{E}_{x\sim P_\text{data}}[\log D(x|c)] + \mathbb{E}_{z\sim P_z}[\log(1-D(G(z|c)))]$$0–$$V(D,G)=\mathbb{E}_{x\sim P_\text{data}}[\log D(x|c)] + \mathbb{E}_{z\sim P_z}[\log(1-D(G(z|c)))]$$1 for CTGAN-MHE, while requiring a fraction of the compute footprint. Contextual BERT similarity is marginally lower due to coarser cluster-based discretization, but overall synthetic joint distribution fidelity and semantic coherence are preserved (Samanta et al., 2024).

5. Downstream Applications and Advanced Use Cases

CTGAN has been deployed to address a range of domain-specific data synthesis challenges. In intrusion detection, CTGAN is used to oversample rare attack classes in low-prevalence regimes, boosting rare-event recall to $$V(D,G)=\mathbb{E}_{x\sim P_\text{data}}[\log D(x|c)] + \mathbb{E}_{z\sim P_z}[\log(1-D(G(z|c)))]$$2 and achieving overall detection accuracy of $$V(D,G)=\mathbb{E}_{x\sim P_\text{data}}[\log D(x|c)] + \mathbb{E}_{z\sim P_z}[\log(1-D(G(z|c)))]$$3 when coupled with additional sampling/cleaning stages (e.g., SMOTEENN) and deep classifiers (Menssouri et al., 9 Feb 2025).

In causal inference, CTGAN-augmented pipelines generate high-dimensional counterfactual trajectories, preserving baseline covariate distributions via the “skeleton injection” mechanism. CTGAN outperforms TVAE and Transformer VAEs on utility and pre-trend causal falsification metrics when augmenting small-$$V(D,G)=\mathbb{E}_{x\sim P_\text{data}}[\log D(x|c)] + \mathbb{E}_{z\sim P_z}[\log(1-D(G(z|c)))]$$4 longitudinal panels for robust difference-in-differences analyses (Grassi et al., 27 Mar 2026).

6. Limitations, Variants, and Trade-offs

Multi-hot encodings for structured word or phrase columns suffer from combinatorial explosion in input dimensionality, causing excessive memory and slow training. One-hot approaches collapse semantic variability and yield only the observed combinations, failing to generalize sensibly. CTG-KrEW’s embedding-cluster-count representation addresses both, but induces a minor trade-off in fine-grained contextual similarity due to discretization (Samanta et al., 2024).

Variants such as Bi-Discriminator CTGAN (BCT-GAN) introduce multiple critics and alternative conditional masking (e.g., sparse Chi-squared masks), yielding richer training gradients, reduced mode collapse, and improved likelihood-fitness and machine learning efficacy metrics, especially on heterogeneous datasets with substantial categorical structure (Esmaeilpour et al., 2021).

Bayesian extensions (e.g., GACTGAN) inject posterior weight uncertainty via Stochastic Weight Averaging–Gaussian (SWAG), enabling better preservation of tabular structure, higher utility, and lower disclosure risk compared to vanilla CTGAN, all with modest computational overhead (Nasution et al., 25 Feb 2026).

Summary and comparative performance table (drawn from (Samanta et al., 2024)):

Method	CPU Time (s) / Memory (MiB)	Entropy	KL Divergence	Pearson ρ	BERT Similarity
CTGAN-1hot	5–30 / 200–250	Low	High	0.5–0.6	Low
CTGAN-MHE	900–1,200 / 600–700	High	Medium	0.75–0.85	High
CTG-KrEW (K=4)	12–35 / 400–450	High	Low	0.92–0.97	Medium–High

7. Principal Findings and Impact

CTGAN and its derivatives offer a theoretically principled and empirically validated solution for tabular data synthesis under complex, multimodal, and semantically structured regimes. Innovations such as CTG-KrEW’s embedding and clustering pipeline unlock high-efficiency, high-fidelity generation of structured content, preserving contextual, frequency, and co-occurrence properties while achieving large reductions in computational resource usage. These capabilities have catalyzed adoption across domains including privacy-preserving behavior modeling, rare event simulation, and robust causal inference (Samanta et al., 2024, Menssouri et al., 9 Feb 2025, Nasution et al., 25 Feb 2026, Grassi et al., 27 Mar 2026).

References:

• (Xu et al., 2019) “Modeling Tabular Data using Conditional GAN”
• (Samanta et al., 2024) “CTG-KrEW: Generating Synthetic Structured Contextually Correlated Content by Conditional Tabular GAN with K-Means Clustering and Efficient Word Embedding”
• (Menssouri et al., 9 Feb 2025) “A Conditional Tabular GAN-Enhanced Intrusion Detection System for Rare Attacks in IoT Networks”
• (Nasution et al., 25 Feb 2026) “Bayesian Generative Adversarial Networks via Gaussian Approximation for Tabular Data Synthesis”
• (Grassi et al., 27 Mar 2026) “Synthesizing the Counterfactual: A CTGAN-Augmented Causal Evaluation of Palliative Care on Spousal Depression”
• (Esmaeilpour et al., 2021) “Bi-Discriminator Class-Conditional Tabular GAN”