Conditional Synthesis

Unlike unconditional generation, conditional synthesis accepts a conditioning vector (e.g., a class label, a specific feature value, or a continuous attribute range) that directs the sampling process. The model learns the conditional probability distribution P(X | Y=y), allowing users to specify "generate a synthetic record where age > 65 and diagnosis = diabetes." This is critical for augmenting underrepresented subgroups in imbalanced datasets without altering the distribution of other features.

The Conditional Tabular GAN (CTGAN) is the foundational architecture for conditional synthesis on structured data. It introduces three key innovations:

• Mode-specific normalization: Handles non-Gaussian, multi-modal distributions common in tabular data
• Conditional vector: A one-hot encoded constraint fed to both generator and discriminator during training
• Training-by-sampling: The condition is randomly sampled from the real data's discrete columns to ensure the generator sees all possible conditions during training, preventing mode collapse on minority classes

Conditional synthesis enables the generation of rare event scenarios that may be underrepresented or entirely absent from historical data. Financial institutions use it to synthesize fraudulent transaction patterns for specific merchant categories; autonomous vehicle teams generate edge-case sensor data (e.g., "pedestrian at night in rain") to harden perception models. This targeted augmentation improves model robustness against tail-risk events without waiting for real-world occurrences.

In sensitive domains like healthcare, certain disease cohorts may have too few records to safely release or share. Conditional synthesis can generate additional synthetic records for a specific diagnosis code while preserving the statistical correlations with other features (lab results, demographics). This balances datasets for downstream model training without exposing real patient records from minority classes, directly supporting fairness-aware synthesis and reducing disparate model performance across subgroups.

Beyond GANs, Denoising Diffusion Probabilistic Models (DDPMs) have been adapted for conditional synthesis. Classifier-free guidance allows a diffusion model to generate samples conditioned on a label by jointly training on conditional and unconditional objectives. During inference, the model interpolates between the two, producing high-fidelity synthetic data that strictly adheres to the specified constraint. This approach excels in generating high-dimensional data like synthetic time series conditioned on a specific trend or seasonality pattern.

When synthesizing relational databases, conditional synthesis must preserve foreign key relationships across tables. The Synthetic Data Vault (SDV) implements conditional synthesis for multi-table settings by first modeling the parent table, then conditioning child table generation on the synthesized parent keys. This ensures that a synthetic order always references a valid synthetic customer, maintaining the referential integrity required for realistic database testing and application development.

A dual-network architecture where a generator produces synthetic samples and a discriminator attempts to distinguish them from real data. In conditional synthesis, both networks receive the conditioning variable as an additional input.

• Adversarial loss drives realism
• Conditional GAN (cGAN) extends the framework for targeted generation
• Prone to mode collapse without architectural safeguards

A generative model that learns a probabilistic latent space representation of the data. Conditional VAEs (CVAEs) inject the conditioning variable into both the encoder and decoder.

• Produces smooth, continuous latent manifolds
• Enables interpolation between specified conditions
• Often combined with GAN discriminators for sharper outputs (VAE-GAN hybrids)

A generative framework that learns to reverse a gradual noising process. Conditional diffusion models guide the denoising trajectory using classifier signals or classifier-free guidance.

• State-of-the-art sample fidelity for images and tabular data
• Classifier-free guidance enables precise condition following
• Iterative denoising provides fine-grained control over output attributes

The degree to which synthetic data preserves the univariate distributions, multivariate correlations, and aggregate statistics of the real data. Conditional synthesis must maintain fidelity within each specified subgroup.

• Measured via column shape similarity and pair-wise trend correlation
• Boundary adherence ensures synthetic values respect real data ranges
• Critical for downstream model utility

The fundamental balancing act between the strength of privacy guarantees and the statistical usefulness of synthetic data. Conditional synthesis adds complexity: privacy must hold within each conditioned subgroup.

• Stricter privacy budgets degrade minority class fidelity
• Differential privacy provides formal bounds but reduces signal
• Requires per-condition evaluation of re-identification risk