Research map / Graph learning / GNN architectures

AutoSGNN: Automatic Propagation Mechanism Discovery for Spectral Graph Neural Networks

AAAI 2025 · Graph learning › GNN architectures · approach: Oversmoothing and heterophily · 9 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AutoSGNN, which uses GPT-3.5-turbo inside an evolutionary loop with three evolutionary prompts to automatically generate spectral GNN propagation mechanisms for homophilic and heterophilic node classification.

Problem

A single spectral GNN struggles to handle both homophilic and heterophilic graphs, and manually designing spectral GNNs per graph type is costly and limited by expert knowledge.

Approach

Combine LLM code generation with evolution strategies (mutation-like E1, crossover-like E2, and DPO-inspired preference prompt C1) over a unified spectral GNN search space (feature fitting, Laplacian regularization, aggregation terms), so the LLM generates per-dataset spectral GNN code scored by validation accuracy.

Key results

• Under sparse 2.5%:2.5%:95% splits, AutoSGNN ranks first (Wilcoxon-Holm) and reaches 85.13±0.26 on PubMed and 92.51±0.36 on Photo, versus 85.07 for NFGNN/GPRGNN on PubMed.
• Under dense 60/20/20 splits, AutoSGNN achieves 91.97±0.35 on PubMed and 95.12±0.37 on Photo, ranking second overall behind NFGNN.
• Searches take roughly 160-176 minutes per dataset on Cora, CiteSeer and PubMed, positioned between gradient-based NAS and evolutionary NAS in time cost.

Measured effects reported in the paper

Surprising findings

• Spectral GNN architectures found for one dataset often fail to transfer well to other graph types, even within heterophilic or homophilic groups.
• Removing the DPO-inspired preference prompt C1 causes the largest drop in the PubMed ablation (83.95 vs 85.13).

Limitations

• Evaluated only on node classification with nine datasets; results on other tasks are not shown.
• Relies on closed-source GPT-3.5-turbo; LLM dependence and API cost are not quantified.
• Discovered architectures transfer poorly across graph types, so search must be rerun per dataset.

Experimental setup

Datasets: Cora CiteSeer PubMed Computers Photo Chameleon Squirrel Texas Cornell

Benchmarks: Cora CiteSeer PubMed Computers Photo Chameleon Squirrel Texas Cornell

Metrics: node classification accuracy Wilcoxon-Holm rank search time (minutes)

Baselines: GTFGNAS F2GNN Genetic-GNN SANE APPNP GPRGNN FAGCN BernNet JacobiConv NFGNN GCN ChebNet

Builds on: BernNet GPRGNN DPO GraphNAS FunSearch-style LLM heuristic discovery

Closest papers by meaning

• Graph Neural Networks Are More Than Filters: Revisiting and Benchmarking from A Spectral Perspective (ICLR 2025 · similarity 0.90)
• SLOG: An Inductive Spectral Graph Neural Network Beyond Polynomial Filter (ICML 2024 · similarity 0.90)
• Homomorphism Expressivity of Spectral Invariant Graph Neural Networks (ICLR 2025 · similarity 0.89)
• On the Expressive Power of Spectral Invariant Graph Neural Networks (ICML 2024 · similarity 0.88)
• Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 2025 · similarity 0.88)
• Asymmetric Learning for Spectral Graph Neural Networks (AAAI 2025 · similarity 0.88)