MAGiC: Attributed Graph Generation via Mixed-type Diffusion and Coarsening

Abele Mălan, Roberto Gheda, Robert Birke, Lydia Y. Chen

Keywords: diffusion model, attributed graph, coarsening

TL;DR: Mixed-type diffusion for generating graphs with rich node attributes paired with invertible coarsening for improved efficiency

Abstract: Many domains, like social and document networks, model relationships as graphs with rich node attributes and large node counts. However, modern graph generators cater to the discreteness of edge connectivity, at the expense of only allowing categorical node labels, and are limited to relatively small graphs, like molecules, due to scalability challenges. To overcome such challenges, we propose MAGiC, a framework that enables graph diffusion with mixed-type node attributes and improves scalability even in unattributed graph scenarios. At the core of MAGiC, a novel mixed-type diffusion joins discrete diffusion for the graph structure with continuous diffusion for node attribute embeddings in a single model. It enables the generation of nodes with rich attributes while maintaining the graph structure quality benefits of discrete diffusion. Alongside it, we propose an invertible coarsening algorithm and a structure-aware attribute encoder that boost scalability, reducing diffusion memory and computation costs. We evaluate MAGiC against baselines combining unattributed graph and tabular generation on three datasets with rich node attributes. Our solution is on average $12.9\times$ better at capturing attribute--structure interaction and $25.2\%$ better at downstream machine learning tasks. Concurrently, we maintain competitive synthesis quality for simple graphs with single categorical node labels. Moreover, MAGiC's coarsening (and attribute encoder) consistently reduces inference time by $2.5\times$ for simple and rich graphs.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 11240