[Paper Review] SPECTRE: Spectral Conditioning Helps to Overcome the Expressivity Limits of One-shot Graph Generators

Karolis Martinkus, Andreas Loukas+2 more|arXiv (Cornell University)|Apr 4, 2022

Advanced Graph Neural Networks4 citations

TL;DR

SPECTRE proposes a novel GAN-based graph generation framework that overcomes the expressivity limitations of one-shot graph generators by first generating the dominant components of the graph Laplacian spectrum and then conditioning graph generation on these spectral features. This spectral conditioning enables control over global graph structure, significantly improving fidelity and novelty while achieving 4-to-170x faster generation than autoregressive models and outperforming state-of-the-art methods on synthetic and real-world graphs.

ABSTRACT

We approach the graph generation problem from a spectral perspective by first generating the dominant parts of the graph Laplacian spectrum and then building a graph matching these eigenvalues and eigenvectors. Spectral conditioning allows for direct modeling of the global and local graph structure and helps to overcome the expressivity and mode collapse issues of one-shot graph generators. Our novel GAN, called SPECTRE, enables the one-shot generation of much larger graphs than previously possible with one-shot models. SPECTRE outperforms state-of-the-art deep autoregressive generators in terms of modeling fidelity, while also avoiding expensive sequential generation and dependence on node ordering. A case in point, in sizable synthetic and real-world graphs SPECTRE achieves a 4-to-170 fold improvement over the best competitor that does not overfit and is 23-to-30 times faster than autoregressive generators.

Motivation & Objective

• To address the expressivity and mode collapse issues in one-shot graph generators, which struggle to model global graph structure due to reliance on local node interactions.
• To enable the generation of larger, more complex graphs than previously possible with one-shot models by leveraging spectral properties of the graph Laplacian.
• To achieve high modeling fidelity and novelty without sequential generation or node-order dependence, improving upon existing GAN and autoregressive approaches.
• To demonstrate that spectral conditioning enables better control over global graph structure while maintaining differentiability and permutation equivariance in the generation process.

Proposed method

• The method first models the top-k eigenvalues and eigenvectors of the graph Laplacian using a spectral generator, capturing global structural properties such as connectivity and clustering.
• These spectral components are then used to condition a second generator that produces node embeddings and constructs the final graph, ensuring spectral consistency.
• The entire pipeline is end-to-end differentiable and permutation-equivariant, preserving structural invariance under node relabeling.
• The generator uses a GNN-based architecture inspired by GG-GAN, with spectral conditioning applied to the node embedding initialization.
• Training is performed via a GAN objective with a PPGN discriminator, and spectral fidelity is optimized using MMD and Wavelet MMD losses on eigenvalues and eigenvectors.
• The model is trained jointly on both spectral and graph generation objectives, enabling direct control over global structure before local refinement.

Experimental results

Research questions

• RQ1 Can spectral conditioning improve the expressivity and structural fidelity of one-shot graph generators?
• RQ2 Does conditioning on the graph Laplacian spectrum help mitigate mode collapse in GAN-based graph generation?
• RQ3 Can one-shot graph generation achieve performance comparable to autoregressive models while being significantly faster and avoiding sequential generation?
• RQ4 To what extent does spectral conditioning enable the generation of larger, more complex graphs than previous one-shot methods?
• RQ5 How does conditioning on real versus generated spectra affect the quality and novelty of generated graphs?

Key findings

• SPECTRE achieves a 4-to-170-fold improvement in MMD ratio over the best-performing baseline on synthetic and real-world graphs, indicating superior modeling fidelity.
• On the QM9 molecule dataset, SPECTRE achieves a validation and uniqueness score of 31.2 and a novelty score of 29.1, outperforming MolGAN* and GG-GAN* variants.
• SPECTRE is 23-to-30 times faster than autoregressive generators while maintaining or exceeding their performance in fidelity and novelty.
• Conditioning on real spectra improves performance, especially when k is large, suggesting that better spectral generators could further enhance results.
• SPECTRE demonstrates robustness to mode collapse, outperforming models like MolGAN* and GG-GAN* that suffer from this issue on QM9.
• Even when generating from imperfect spectra, SPECTRE produces high-quality graphs, indicating strong generalization and error correction capability in spectral conditioning.