CTGAN

A novel data transformation that handles non-Gaussian and multi-modal continuous columns. Instead of a single min-max or Z-score normalization, CTGAN uses a variational Gaussian mixture model (VGM) to estimate the number of modes and fit a distribution for each column.

• Each value is represented as a one-hot vector indicating its sampled mode and a scalar representing the normalized value within that mode.
• This prevents the gradient-based generator from being overwhelmed by complex, sparse distributions.
• Crucial for columns like transaction amounts or claim values that exhibit heavy-tailed or clustered behavior.

A mechanism to combat mode collapse and ensure the generator faithfully represents all discrete categories, including rare classes.

• A conditional vector is constructed by randomly selecting a discrete column and one of its categories.
• The generator is conditioned on this vector, forcing it to produce a sample belonging to that specific category.
• The training data is sampled uniformly across all categories, not proportionally to their frequency, ensuring minority classes are equally represented during training.
• This is critical for preserving the statistical integrity of imbalanced columns like fraud flags or disease codes.

CTGAN employs deep fully connected (dense) neural networks with residual connections for both the generator and discriminator, optimized for tabular data's unique structure.

• The generator transforms a random noise vector and a conditional vector into a synthetic row.
• The discriminator uses techniques like pac (packing) , where multiple samples are concatenated before being passed to the discriminator, which helps prevent mode collapse by providing a broader view of the data distribution.
• This architecture is specifically designed to capture the complex, non-linear inter-column correlations that simpler models miss.

CTGAN is architected from the ground up to handle the heterogeneous mix of data types found in real databases without manual feature engineering.

• Continuous columns: Processed via mode-specific normalization.
• Discrete columns: Represented as one-hot encoded vectors.
• Categorical columns: Treated as discrete and balanced via conditional training.
• The output layer uses a tanh activation for continuous values (scaled back to the original range) and gumbel-softmax activations for discrete columns, allowing for end-to-end differentiable training across all data types simultaneously.

CTGAN leverages the Wasserstein distance with a gradient penalty as its loss function, a significant improvement over the standard Jensen-Shannon divergence used in vanilla GANs.

• Provides more stable and meaningful gradients, even when the real and generated distributions have non-overlapping support.
• The gradient penalty term enforces a 1-Lipschitz constraint on the discriminator (critic), preventing exploding or vanishing gradients.
• This results in far more reliable convergence and higher-fidelity synthetic data, especially for complex, high-dimensional tables.

CTGAN is a core model within the Synthetic Data Vault (SDV) ecosystem, providing a production-ready pipeline from raw data to a deployed synthetic dataset.

• Metadata API: Automatically infers column types, constraints, and relationships from a real table.
• Quality Evaluation: Integrates directly with SDMetrics to generate a comprehensive diagnostic report measuring statistical fidelity, column shapes, and pair-wise trend preservation.
• Constraint Preservation: Can be combined with SDV's constraint framework to enforce business logic, like ensuring a start_date is always before an end_date in the synthetic data.

A critical failure condition in GAN training where the generator learns to produce only a limited variety of outputs, ignoring entire regions of the real data distribution. In tabular contexts, this manifests as synthetic data that fails to represent minority classes or rare value combinations. CTGAN combats mode collapse through:

• Mode-specific normalization: Handling non-Gaussian, multi-modal distributions per column
• Conditional generation: Explicitly training the generator to sample from all discrete categories
• Training-by-sampling: Ensuring balanced exposure to all values during training

The degree to which synthetic data preserves the statistical properties of the original dataset. High-fidelity synthetic data must maintain:

• Univariate distributions: Column-level shapes, means, variances, and quantiles
• Multivariate correlations: Pairwise relationships and higher-order interactions between columns
• Boundary adherence: Respecting min/max ranges and domain constraints
• Category balance: Preserving the frequency distribution of discrete values CTGAN's mode-specific normalization and conditional vector mechanisms are explicitly designed to maximize fidelity across these dimensions.