AutoSGNN: Automatic Propagation Mechanism Discovery for Spectral Graph Neural Networks

TL;DR

AutoSGNN tackles the limitation of single spectral GNNs across diverse graph types by automatically discovering propagation mechanisms with a framework that combines large language models and evolutionary strategies. It unifies the spectral GNN search space and continuously evolves both the design ideas and code, enabling parallel generation and evaluation. Across nine datasets, AutoSGNN consistently rivals or surpasses state-of-the-art spectral GNNs and GNN-NAS methods while offering favorable time efficiency. The work demonstrates the viability of LLM-assisted propagation mechanism generation for graph learning and highlights potential for dataset-type-aware auto-design of graph foundation models.

Abstract

In real-world applications, spectral Graph Neural Networks (GNNs) are powerful tools for processing diverse types of graphs. However, a single GNN often struggles to handle different graph types-such as homogeneous and heterogeneous graphs-simultaneously. This challenge has led to the manual design of GNNs tailored to specific graph types, but these approaches are limited by the high cost of labor and the constraints of expert knowledge, which cannot keep up with the rapid growth of graph data. To overcome these challenges, we propose AutoSGNN, an automated framework for discovering propagation mechanisms in spectral GNNs. AutoSGNN unifies the search space for spectral GNNs by integrating large language models with evolutionary strategies to automatically generate architectures that adapt to various graph types. Extensive experiments on nine widely-used datasets, encompassing both homophilic and heterophilic graphs, demonstrate that AutoSGNN outperforms state-of-the-art spectral GNNs and graph neural architecture search methods in both performance and efficiency.

Figures (13)

• Figure 1: Radar chart of AutoSGNN and other spectral GNNs on 9 datasets.
• Figure 2: The overview of AutoSGNN. In this process, the LLM is used to generate population individuals, while other components are responsible for evaluation, selection, and describing requirements to optimize the LLM preferences, thus achieving the overall optimization process.
• Figure 3: Individual Content Schematic.
• Figure 4: Prompt Strategies for LLMs.
• Figure 5: The iterative search process of AutoSGNN with different Elite Individual numbers for node classification experiment on the PubMed dataset. Each dot represents the node classification result of a searched condidate individual on the PubMed dataset. The dashed line represents the overall convergence trend of elite individuals.