CTGAN vs TVAE: Tabular Data Synthesis

CTGAN excels at modeling complex, non-Gaussian numerical distributions column by column because its conditional GAN architecture uses a mode-specific normalization technique. For example, on a financial dataset with heavy-tailed transaction amounts, CTGAN can preserve the exact min/max boundaries and multi-modal peaks, often achieving a Kolmogorov-Smirnov (KS) complement score above 0.90 for continuous variables. This makes it a powerful tool when the statistical fidelity of individual columns is the primary success metric.

TVAE takes a fundamentally different approach by learning a compressed, continuous latent representation of entire rows using a variational autoencoder. This results in a smoother, more stable training process that naturally handles missing values and captures inter-row relationships. The key trade-off is that while TVAE better preserves the overall covariance structure between columns, it may slightly oversmooth the extreme tails of a distribution, producing fewer realistic outliers than a GAN-based method.

The key trade-off: If your priority is maximizing the realism of individual column distributions, especially for skewed numericals and high-cardinality categoricals, choose CTGAN. If you prioritize training stability, faster convergence, and the preservation of multi-column correlations for downstream machine learning tasks, TVAE is the more robust choice. For regulated use cases, consider CTGAN when outlier fidelity is critical for risk modeling, and TVAE when a stable, general-purpose synthetic twin is needed for application testing.

CTGAN excels at modeling complex, non-Gaussian numerical distributions and high-cardinality categorical features because its mode-specific normalization and conditional generator directly address multi-modal data. For example, in benchmarks on the Adult Income dataset, CTGAN achieves a higher column-wise correlation preservation score (0.85 vs. 0.78) and better captures rare category combinations, making it the superior choice for applications like fraud detection where long-tail fidelity is critical.

TVAE takes a fundamentally different approach by learning a compressed, continuous latent representation of entire rows, which results in more stable and convergent training. This architecture avoids the adversarial game's mode collapse risk entirely. In regulated industry use cases, TVAE often demonstrates a lower nearest neighbor distance ratio (NNDR), suggesting it provides stronger empirical privacy by being less likely to memorize and regurgitate exact training outliers, a key advantage for HIPAA or GDPR compliance.

The key trade-off: If your priority is maximizing statistical fidelity, especially for multi-modal numerical columns and preserving complex inter-feature correlations, choose CTGAN. If you prioritize training stability, a principled latent space for data exploration, and a potentially lower risk of overfitting to sensitive records, choose TVAE. For a direct comparison of their privacy risk profiles under formal guarantees, see our analysis of DP-GAN vs DP-VAE.