SDV Methods

PETsARD integrates the SDV (Synthetic Data Vault) package, providing various advanced synthetic data generation algorithms.

⚠️

Important Notice: The built-in SDV integration is planned for deprecation in the future. We recommend using SDV Custom Methods for better flexibility and long-term support.

ℹ️

Note: This document only provides YAML configuration examples, without Jupyter notebook examples.

Usage Examples

Loader:
  load_benchmark_with_schema:
    filepath: benchmark://adult-income
    schema: benchmark://adult-income_schema

Synthesizer:
  gaussian:
    method: sdv-single_table-gaussiancopula

  ctgan:
    method: sdv-single_table-ctgan

  copulagan:
    method: sdv-single_table-copulagan

  tvae:
    method: sdv-single_table-tvae

Methods Overview

Method	method Setting	Features	GPU
GaussianCopula	`default` or `sdv-single_table-gaussiancopula`	Fast, suitable for large data	✗
CTGAN	`sdv-single_table-ctgan`	High quality, complex patterns	✓
CopulaGAN	`sdv-single_table-copulagan`	Balances statistics & deep learning	✓
TVAE	`sdv-single_table-tvae`	Stable training, fast convergence	✓

Method Details

GaussianCopula

Classical statistical distribution-based method, fast execution suitable for quick prototyping.

Features:

✓ Fast, suitable for large data
✓ Low computational requirements
✗ Primarily captures linear correlations

CTGAN

GAN-based deep learning method with best generation quality.

Features:

✓ High-quality synthetic data
✓ Suitable for complex patterns
✗ Longer training time

Default Parameters:

epochs: 300
batch_size: 500
generator_lr: 0.0002
discriminator_lr: 0.0002

CopulaGAN

Combines Copula statistics with GAN, suitable for mixed-type data.

Features:

✓ Balances statistics & deep learning
✓ Better marginal distribution simulation
✓ Suitable for continuous & discrete mixed data

Default Parameters:

epochs: 300
batch_size: 500
default_distribution: beta

TVAE

VAE-based generative model with stable training process.

Features:

✓ Stable training process
✓ Better convergence
✓ Suitable for medium-scale data

Default Parameters:

epochs: 300
batch_size: 500
encoder_layers: [128, 128]
decoder_layers: [128, 128]

Automatic Features

Schema Conversion

PETsARD automatically converts internal Schema to SDV Metadata

Automatic Parameters

All methods automatically enable:

enforce_rounding: Integer rounding
enforce_min_max_values: Value range enforcement (GaussianCopula, TVAE)

GPU Detection

Deep learning methods (CTGAN, CopulaGAN, TVAE) automatically detect and use GPU.

Selection Guide

Scenario	Recommended Method
Quick testing	GaussianCopula
High quality needs	CTGAN
Mixed-type data	CopulaGAN
Medium data	TVAE
Large data	GaussianCopula

Available Distributions

GaussianCopula and CopulaGAN support:

norm: Normal distribution
truncnorm: Truncated normal distribution (default)
beta: Beta distribution
gamma: Gamma distribution
uniform: Uniform distribution
gaussian_kde: Kernel density estimation

Limitations

Built-in Integration Limits

✗ Cannot adjust training parameters (epochs, batch_size, etc.)
✗ Cannot specify distribution types
✗ Cannot manually select CPU/GPU

Important Notes

Deep learning methods train faster on GPU
Default 300 epochs, CPU training may be time-consuming
Large datasets with deep learning require significant memory
Built-in integration uses fixed parameters, cannot be adjusted

Custom Method: SDV