Synthcity is a comprehensive Python library for generating and evaluating synthetic tabular, time-series, survival, and image data. It targets researchers and practitioners needing to augment datasets, ensure privacy, or test fairness, offering a wide array of state-of-the-art generative models and robust evaluation metrics.

How It Works

Synthcity employs a plugin-based architecture, allowing easy integration of diverse generative models including GANs (AdsGAN, CTGAN, PATEGAN), VAEs (TVAE, RTVAE), Normalizing Flows, Bayesian Networks, and LLM-based models (GReaT). It supports specialized generators for time-series and survival data, alongside privacy-focused methods like DECAF and DP-GAN. The library also provides a rich suite of evaluation metrics for data quality, privacy, and fairness.

Quick Start & Requirements

• Install via pip: pip install synthcity
• Requires Python 3.8+.
• Official documentation: https://synthcity.readthedocs.io/en/latest/
• Colab notebook: https://colab.research.google.com/drive/1Vr2PJswgfFYBkJCm3hhVkuH-9dXnHeYV?usp=sharing

Highlighted Details

• Supports tabular, time-series, survival analysis, and image data generation.
• Integrates numerous generative models, including GANs, VAEs, Normalizing Flows, Bayesian Networks, and LLM-based models.
• Offers extensive evaluation metrics for data quality, privacy (k-anonymity, l-diversity), and fairness.
• Includes specialized generators for time-series (TimeGAN, FourierFlows) and survival analysis (SurvivalGAN).

Maintenance & Community

• Active development by van der Schaar Lab.
• Slack community available: https://join.slack.com/t/vanderschaarlab/shared_invite/zt-1pzy8z7ti-zVsUPHAKTgCd1UoY8XtTEw

Licensing & Compatibility

• License: Apache 2.0.
• Compatible with commercial and closed-source applications.

Limitations & Caveats

• Does not handle missing data; imputation is required beforehand (e.g., using HyperImpute).
• Image generation architectures are noted as not state-of-the-art.