Relational Data Generation with Graph Neural Networks and Latent Diffusion Models

This repository contains the code for the paper "Relational Data Generation with Graph Neural Networks and Latent Diffusion Models"

This project presents a novel method for synthetic relational data generation that uses a graph representation of a relational database. We combine the expressive power of graph neural networks and generative capabilities of latent diffusion models to achieve state-of-the-art performance with respect to multi-table fidelity on the SyntheRela benchmark.

Installation

Create environment and install requirements

conda create -n relgdiff python=3.10
conda activate relgdiff
pip install -r requirements.txt

Example

Download the datasets from the SyntheRela benchmark here.

Preprocess data

python src/scripts/preprocess_data.py --dataset-name DATASET_NAME

Training

python src/scripts/train.py --dataset-name DATASET_NAME

Sampling

python src/scripts/sample.py --dataset-name DATASET_NAME 

Citing

@inproceedings{
  hudovernik2024relational,
  title={Relational Data Generation with Graph Neural Networks and Latent Diffusion Models},
  author={Valter Hudovernik},
  booktitle={NeurIPS 2024 Third Table Representation Learning Workshop},
  year={2024},
  url={https://openreview.net/forum?id=MNLR2NYN2Z}
}