Synthcity

Synthcity abstracts away the implementation details of dozens of generative models behind a single, consistent API. This allows data scientists to swap between Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and Diffusion Models without rewriting their data loading or evaluation pipelines.

• Consistent Interface: All models expose standard fit() and generate() methods.
• Extensible Design: New models can be added as plugins without modifying the core library.
• Broad Model Support: Includes CTGAN, TVAE, PATEGAN, ADS-GAN, and many more.

Synthcity integrates a robust set of metrics to quantify the statistical fidelity, privacy risk, and downstream utility of generated synthetic data. This moves evaluation beyond visual inspection to rigorous, reproducible benchmarks.

• Fidelity Metrics: Measures how well the synthetic data preserves column shapes, pair-wise correlations, and boundary adherence.
• Privacy Metrics: Quantifies re-identification risk, membership inference attack susceptibility, and attribute inference leakage.
• Utility Metrics: Implements the Train-Synthetic-Test-Real (TSTR) paradigm to assess how well models trained on synthetic data perform on real test sets.

The library includes specialized generators designed explicitly to enforce formal privacy guarantees, such as Differential Privacy. These models inject calibrated noise during training to bound the information leakage about any single training record.

• DP-GAN: A differentially private GAN implementation that adds noise to gradients during training.
• PATE-GAN: Uses the Private Aggregation of Teacher Ensembles framework to train a student generator with provable privacy.
• Privacy Budget Tracking: Allows users to configure and monitor the epsilon parameter, directly managing the privacy-utility trade-off.

Beyond static tabular data, Synthcity supports the generation of complex sequential data and allows for fine-grained control over the output distribution through conditional synthesis.

• Time-Series Models: Dedicated generators like TimeGAN capture temporal dynamics, seasonality, and autocorrelations for synthetic time series.
• Conditional Generation: Users can specify constraints to generate data for a specific class or subgroup, enabling targeted data amplification for imbalanced datasets.
• Mixed Data Types: Handles categorical, continuous, integer, and date features natively.

Synthcity is built as a benchmarking framework first. It provides standardized datasets, predefined experimental workflows, and leaderboards to ensure that comparisons between different generative models are fair and reproducible.

• Standardized Datasets: Includes a curated set of canonical benchmarks for evaluating generative models.
• Reproducible Workflows: Experiments are defined declaratively, ensuring that results can be replicated exactly.
• Model Comparison: Directly compare the fidelity, privacy, and utility scores of multiple generators side-by-side.

The library's flexibility makes it suitable for generating high-fidelity synthetic data across regulated industries. It enables the creation of synthetic health data and synthetic financial data that preserve complex statistical properties without exposing sensitive records.

• Healthcare: Generate patient trajectories and clinical records that maintain physiological correlations while ensuring patient privacy.
• Finance: Create transaction logs and market data that replicate fraud signatures and temporal patterns for robust model development.
• Fairness-Aware Synthesis: Includes methods to generate data that corrects for historical biases, promoting demographic parity.