Data-Centric-AI-Community/fg-data-synthetic

Synthetic data generators for tabular and time-series data

What it solves

This project provides a way to generate synthetic tabular and time-series data. This is useful for ensuring privacy compliance when sharing data, removing bias, balancing datasets, and augmenting existing datasets without using real-world identifiable information.

How it works

The library leverages various generative models to replicate the statistical properties of real data. For tabular data, it uses architectures like CTGAN, CGAN, WGAN, and Gaussian Mixture models (which allow for fast generation without a GPU). For sequential or time-series data, it uses models like TimeGAN and DoppelGANger. All deep learning models are implemented using TensorFlow 2.0.

Who it’s for

Data scientists and machine learning engineers who need high-quality synthetic datasets for development, testing, or privacy-preserving data sharing.

Highlights

  • Low-code UI: A Streamlit-based application that allows users to train synthesizer models and generate/profile synthetic data samples via a user interface.
  • Diverse Model Support: Includes a wide range of GAN architectures (e.g., CTGAN, WGAN-GP, DRAGAN) and Gaussian Mixture models.
  • GPU-optional: The Gaussian Mixture model provides a fast start for synthetic data generation without requiring a GPU.
  • Time-Series Capability: Dedicated support for sequential data generation using TimeGAN and DoppelGANger.