Conditional Tabular GAN (CTGAN)

CTGAN addresses non-Gaussian and multimodal distributions in continuous columns by using a variational Gaussian mixture model (VGM). Instead of a single min-max scaler, each continuous value is represented as a one-hot vector indicating the sampled mode and a scalar representing the normalized value within that mode. This prevents the generator from collapsing to simplistic unimodal approximations and accurately captures complex marginal distributions like those found in financial transactions or actuarial tables.

To combat mode collapse on discrete columns with severe class imbalance, CTGAN introduces a conditional vector. During training, a discrete column and a specific category are randomly selected. The conditional vector is constructed as a one-hot encoding of this condition and fed to the generator. The generator is then penalized if it fails to produce a sample matching the specified condition. This training-by-sampling mechanism forces the model to explore all minority classes, ensuring the synthetic data faithfully represents rare but critical events like fraud or system failures.

CTGAN employs deep fully connected (dense) neural networks with batch normalization and leaky ReLU activations for both the generator and discriminator. The generator transforms a latent noise vector concatenated with the conditional vector into a synthetic row. The discriminator evaluates the authenticity of the generated row against real data. This architecture is designed to capture the complex inter-attribute correlations inherent in relational tables, such as the dependency between age, income, and credit score, without assuming column independence.

CTGAN integrates the PacGAN (Packing) framework into its discriminator to enhance stability and prevent mode collapse. Instead of classifying single samples as real or fake, the discriminator processes a pack of multiple samples (default: 10) jointly. This forces the discriminator to evaluate the joint distribution of the batch, making it significantly harder for the generator to trick the network by simply memorizing a few high-probability modes. The result is a more robust gradient signal and higher-fidelity synthetic data.

To enable end-to-end differentiable training over discrete columns, CTGAN applies the Gumbel-Softmax reparameterization trick during generation. Instead of outputting hard argmax decisions, the generator produces a softmax distribution over categories, and the Gumbel-Softmax provides a continuous, differentiable approximation of sampling from this distribution. This allows gradients to flow through the discrete sampling process back to the generator, enabling standard backpropagation to optimize the generation of categorical features like country codes or product IDs.

CTGAN automatically detects and handles mixed data types without manual feature engineering. It classifies columns into continuous (floats, integers with many unique values) and discrete (categorical strings, integers with few unique values). Continuous columns are processed with the VGM, while discrete columns are one-hot encoded. This automated pipeline allows CTGAN to be applied directly to raw enterprise tables containing a mix of demographic, transactional, and behavioral data without requiring domain experts to specify distributional assumptions.

A preprocessing technique critical to CTGAN that handles non-Gaussian and multimodal distributions in continuous columns. Unlike simple min-max scaling, it uses a variational Gaussian mixture model to estimate the number of modes and represent each value as a one-hot vector indicating its mode and a scalar representing its normalized position within that mode. This prevents the generator from smoothing out distinct peaks in the data, ensuring high statistical fidelity for columns like income or claim amounts.

The conditional training logic that solves the class imbalance problem in tabular data. During training, CTGAN randomly selects a discrete column and a specific category, then forces the generator to produce a row matching that condition. This ensures that minority classes are adequately represented in the synthetic data, preventing mode collapse where the model ignores rare but critical categories like specific disease codes or fraud flags.

The adversarial training framework often integrated with CTGAN to improve training stability. Instead of the standard discriminator loss, it uses the Wasserstein distance to measure the difference between real and generated distributions. The gradient penalty enforces a Lipschitz constraint, preventing exploding gradients and providing a meaningful loss curve that correlates with sample quality, making it easier to diagnose training convergence.

The process of converting mixed data types into a dense numerical matrix suitable for neural networks. CTGAN applies distinct strategies:

• Discrete columns: Represented as one-hot vectors.
• Continuous columns: Represented using mode-specific normalization. This hybrid encoding preserves the semantic meaning of categorical variables while accurately capturing the complex distributions of numerical features, forming the foundation of the generator's input representation.

A formal privacy layer that can be added to CTGAN training by clipping gradients and injecting calibrated Gaussian noise during optimization. This transforms the generator into a differentially private mechanism, providing a mathematical guarantee (epsilon) that the synthetic data does not memorize individual training records. This is essential for releasing synthetic data that legally qualifies as anonymized under regulations like GDPR.