synthetic-tabular-data-generation

Synthetic Tabular Data Generation Skill Guide

This skill governs the training, sampling, validation, and privacy auditing of synthetic tabular datasets using Conditional GANs and Copula models.

1. Generative Modeling Pipeline for Tabular Data

Tabular data presents mixed continuous and discrete columns with non-Gaussian, multi-modal distributions.

[ Original Sensitive Dataset ]
               |
               v
[ Data Preprocessing & Metadata Declaration ]
  |-- Discrete / Categorical Encoding (One-hot, Embedding)
  |-- Continuous Normalization via Variational Gaussian Mixture Models (VGM)
               |
               v
[ Generative Architecture Training ]
  |-- CTGAN (Conditional WGAN-GP with Generator & Discriminator)
  |-- TVAE (Tabular Variational Autoencoder)
               |
               v
[ Synthetic Sample Generation & Evaluation ]
  |-- Fidelity Evaluation (Marginal distribution match, Correlation matrices)
  |-- Privacy Audit (Distance to Closest Record [DCR], Membership leakage)

2. Production Code Implementations

A. CTGAN Model Training & Synthetic Sampling (Python / SDV)

import pandas as pd
from sdv.metadata import SingleTableMetadata
from sdv.single_table import CTGANSynthesizer
from sdv.evaluation.single_table import evaluate_quality, run_diagnostic


def generate_synthetic_dataset(
    real_data: pd.DataFrame,
    primary_key_col: str,
    epochs: int = 300,
    num_samples: int = 10000,
) -> pd.DataFrame:
    # 1. Infer metadata schema
    metadata = SingleTableMetadata()
    metadata.detect_from_dataframe(data=real_data)
    metadata.set_primary_key(column_name=primary_key_col)

    # 2. Configure CTGAN Synthesizer
    synthesizer = CTGANSynthesizer(
        metadata,
        epochs=epochs,
        batch_size=500,
        verbose=True,
    )

    # 3. Train on sensitive dataset
    print("[Synthetic] Training CTGAN Synthesizer...")
    synthesizer.fit(real_data)

    # 4. Generate synthetic records
    synthetic_data = synthesizer.sample(num_rows=num_samples)

    # 5. Evaluate Data Quality
    quality_report = evaluate_quality(
        real_data=real_data,
        synthetic_data=synthetic_data,
        metadata=metadata,
    )
    print(f"[Quality] Overall Quality Score: {quality_report.get_score():.2%}")

    return synthetic_data