RelDiff: Relational Data Generative Modeling with Graph-Based Diffusion Models

This repository provides the official implementation of the paper "RelDiff: Relational Data Generative Modeling with Graph-Based Diffusion Models".

Latest Update

• [2025.]：Our code is at the final stage of cleaning up. Please check back soon for its release!

Introduction

Figure 1: A high-level overview of RelDiff

RelDiff is a novel generative framework for synthesizing relational databases with arbitrarily complex schemas, achieving high fidelity and utility. Its key innovations include:

1. A principled framework for generating foreign key structures in relational databases, incorporating hard constraints for referential integrity via Bayesian stochastic block models.
2. A joint diffusion model for synthesizing mixed-type attributes, utilizing GNNs to capture global inter-table dependencies.
3. Explicitly modeling dimension tables as a distinct data type and defining our diffusion model in data space.

The schema of RelDiff is presented in the figure above.