RelGen

RelGen is the abbreviation of Relation Generation. This tool is used to generate relational data in databases. On the other hand, the pronunciation of Rel is similar to Real, which means that the relational data generated by RelGen is very real.

Overview

RelGen is a Python library designed to generate real relational data for users. RelGen uses a variety of advanced deep generative models and algorithms to learn data distribution from real data and generate high-quality simulation data.

Figure: RelGen Overall Architecture

Features

Generate relational data using deep generative models. RelGen provides a variety of deep generative models, including Generative Adversarial Network (GAN), Autoregressive Model (AR Model) and Diffusion Model.

Generate data for multiple relational tables. RelGen can flexibly generate data for multiple relational tables in the database, so that the data distribution of each generated table is close to that of the original table, and the joined table can also be similar to the original joined table.

Evaluate the quality of generated relational data. RelGen evaluates the generated relational data in terms of fidelity, privacy, and diversity, and visualizes the quality of the generated relational data using histogram and t-SNE plot.

Installation

RelGen requires Python version 3.7 or later.

Install from pip

pip install relgen

Install from source

git clone https://github.com/ruc-datalab/RelGen.git && cd RelGen
pip install -r requirements.txt

Quick-Start

Load Dataset

Load a demo dataset to get started. This dataset is a single table describing the census.

Load metadata for the census dataset.

from relgen.data.metadata import Metadata

metadata = Metadata()
metadata.load_from_json("datasets/census/metadata.json")

Load data for the census dataset.

import pandas as pd

data = {
    "census": pd.read_csv("datasets/census/census.csv")
}

Encapsulate the census dataset and process it.

from relgen.data.dataset import Dataset

dataset = Dataset(metadata)
dataset.fit(data)

Generating Data

Train the synthesizer.

from relgen.synthesizer.arsynthesizer import MADESynthesizer

synthesizer = MADESynthesizer(dataset)
synthesizer.fit(data)

Generate relational data.

sampled_data = synthesizer.sample()

Evaluating Data

Compare real data and generated data to evaluate the quality of generated data.

from relgen.evaluator import Evaluator

evaluator = Evaluator(data["census"], sampled_data["census"])

Show comparison histogram of data distribution between real data and generated data.

evaluator.eval_histogram(columns=["age", "sex", "relationship"])

Show comparison t-SNE plot of data distribution between real data and generated data.

evaluator.eval_tsne()