Research map / Graph learning / GNN architectures

SGNNBench: A Holistic Evaluation of Spiking Graph Neural Networks on Large-scale Graphs

NeurIPS 2026 · Graph learning › GNN architectures · approach: Simple and scalable GNNs

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.

Benchmarks 9 spiking GNNs against classic full-precision GNNs on 18 static and dynamic graph datasets, analyzing effectiveness, energy/memory efficiency, and architectural components.

Problem

Existing spiking graph neural network (SGNN) studies often use unfair or outdated baselines and lack systematic evaluation of effectiveness, energy efficiency, and design choices on large-scale graphs.

Approach

A unified benchmark that re-implements and tunes 9 SGNNs under a common hyperparameter search, compares them with tuned classic GNNs, measures theoretical energy using a firing-rate-based FLOPs formula, and ablates spiking neurons, normalization, and skip/jumping-knowledge components.

Key results

• SGNNs reach roughly 92%-99% of the best full-precision GNN performance on static homophily graphs, with an average gap of about 2.69% on 9 datasets.
• SiGNN outperforms TGAT on dynamic graphs, with Micro-F1 gains of 3.51%, 4.20%, 4.55% at 40%, 60%, 80% training ratios on DBLP-like comparisons reported in the paper.
• SpikingGCN and SpikeGCL achieve about 29.91x and 27.63x lower average theoretical energy than full-precision SGC on four datasets.

Surprising findings

• Adding multi-step and jumping-knowledge components substantially improves SGNNs on heterophily graphs, e.g., GC-SNN with JK gains 7.83% on Roman-Empire.
• SiGNN, despite high accuracy, consumes more energy than most full-precision GNNs due to retaining results from all previous timesteps.

Limitations

• Several SGNN baselines (GC-SNN, GA-SNN) are unofficial re-implementations since source code was unavailable.
• Temporal-coding SGNNs designed for knowledge graph completion are excluded from the benchmark.
• Energy is estimated theoretically from FLOPs and firing rates rather than measured on neuromorphic hardware.

Experimental setup

Datasets: Cora CiteSeer PubMed Computers Photo CS Physics ogbn-proteins ogbn-arxiv Squirrel Chameleon Roman-Empire

Benchmarks: Cora CiteSeer PubMed Computers Photo CS Physics ogbn-proteins ogbn-arxiv Squirrel Chameleon Roman-Empire

Metrics: accuracy ROC-AUC Macro-F1 Micro-F1 model size (MB) peak memory (GB) theoretical energy (mJ)

Baselines: GCN SAGE GAT SGC EvolveGCN TGAT ROLAND GC-SNN GA-SNN SpikingGCN SpikeNet SpikeGCL

Builds on: GCN SGC LIF neuron PLIF Jumping Knowledge SpikeNet SiGNN SpikeGCL

Closest papers by meaning

• A Graph is Worth 1-bit Spikes: When Graph Contrastive Learning Meets Spiking Neural Networks (ICLR 2024 · similarity 0.89)
• Delay-DSGN: A Dynamic Spiking Graph Neural Network with Delay Mechanisms for Evolving Graph (ICML 2025 · similarity 0.89)
• Temporal Spiking Neural Networks with Synaptic Delay for Graph Reasoning (ICML 2024 · similarity 0.88)
• Dynamic Spiking Graph Neural Networks (AAAI 2024 · similarity 0.88)
• Spiking Graph Neural Network on Riemannian Manifolds (NeurIPS 2024 · similarity 0.88)
• Graph Neural Networks Are More Than Filters: Revisiting and Benchmarking from A Spectral Perspective (ICLR 2025 · similarity 0.87)