Are Expressive Encoders Necessary for Discrete Graph Generation?

TL;DR

This paper introduces GenGNN, a modular message-passing framework for graph generation that achieves high validity and faster inference, questioning the necessity of highly expressive encoders like transformers.

Contribution

The paper presents GenGNN, a new GNN backbone for discrete graph generation, demonstrating competitive validity and efficiency, and provides insights into GNN expressiveness in diffusion models.

Findings

GenGNN achieves over 90% validity on Tree and Planar datasets.

GenGNN with diffusion models is 2-5x faster in inference.

Residual connections are crucial to prevent oversmoothing in complex graphs.

Abstract

Discrete graph generation has emerged as a powerful paradigm for modeling graph data, often relying on highly expressive neural backbones such as transformers or higher-order architectures. We revisit this design choice by introducing GenGNN, a modular message-passing framework for graph generation. Diffusion models with GenGNN achieve more than 90% validity on Tree and Planar datasets, within margins of graph transformers, at 2-5x faster inference speed. For molecule generation, DiGress with a GenGNN backbone achieves 99.49% Validity. A systematic ablation study shows the benefit provided by each GenGNN component, indicating the need for residual connections to mitigate oversmoothing on complicated graph-structure. Through scaling analyses, we apply a principled metric-space view to investigate learned diffusion representations and uncover whether GNNs can be expressive neural backbones for discrete diffusion.