DLGrapher: Dual Latent Diffusion for Attributed Graph Generation

Abele Mălan, Robert Birke, Lydia Y. Chen

Keywords: graph generation, diffusion model, attributed graph

TL;DR: We propose DLGrapher the first dual diffusion-based synthesizer for large graphs attributed with rich node features.

Abstract:

Graphs for applications like social data and financial transactions are particularly complex, with large node counts and high-dimensional features. State-of-the-art diffusion graph synthesizers model the node structure via discrete diffusion and are, unfortunately, limited to small-scale graphs with few to no features. In contrast, continuous diffusion models capture rich node features well, but have issues faithfully modelling connectivity. In this paper, we design DLGrapher, a dual latent diffusion framework for jointly synthesizing large graph structures and high-dimension node features. DLGrapher models node features and structure as a joint latent representation. Structure-wise, we design a reversible coarsening scheme to merge pairs of similar neighboring nodes and their respective edges after encoding node features through a structure-aware variational autoencoder. To capture the dependencies between node features and the graph structure, DLGrapher trains a single diffusion over a dual denoising objective, one for the continuous node representations and another for the discrete edge connectivity. We extensively evaluate DLGrapher's performance on three complex social graph datasets against baselines combining tabular and graph synthesizers. Our solution fares 12.9x better at statistically capturing feature-structure interaction and 25.2% better at downstream tasks thanks to the dual diffusion on average and the latent compressed representation increases throughput by 2.5X. Furthermore, we maintain competitive synthesis quality for simple-featured molecular graphs and structure-only synthetic graphs while drastically reducing computation in the latter case.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8287