Reproduce “RelDiff: Relational Data Generative Modeling with Graph-Based Diffusion Models” on OpenResearch

About this paper

Valter Hudovernik, Minkai Xu, Juntong Shi, Lovro Šubelj, Stefano Ermon, Erik Štrumbelj, Jure Leskovec

Developed by researchers at the University of Ljubljana and Stanford University, RELDIFF introduces a generative model for relational databases that explicitly models the foreign key graph structure using a 2K+SBM generator and performs joint attribute synthesis via a graph-conditioned diffusion process. This approach achieved a 25.3% average improvement in k-hop correlations on the ClavaDDPM benchmark and demonstrated a 7x lower performance degradation from single-table to multi-table fidelity compared to prior methods.

Key insights

• Decoupling the generation of relational database structure from attribute synthesis allows for more principled and accurate modeling of each component.
• Explicitly modeling the foreign key graph with a 2K+SBM provides a robust method for preserving structural fidelity, including exact cardinalities and referential integrity, which are critical for relational data.
• A joint, graph-conditioned diffusion process using Graph Neural Networks effectively captures both intra-table and complex inter-table dependencies simultaneously, overcoming limitations of sequential or conditional single-table generation.

Summary by alphaXiv.