Research map / Graph learning / GNN architectures
Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence
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.
Re-evaluates classic GNNs (GCN, GIN, GatedGCN) enhanced with six standard techniques (GNN+) on 14 graph-level datasets and finds they match or beat graph transformers while running faster.
Problem
Whether classic message-passing GNNs can match graph transformers on graph-level tasks (graph classification and regression), which the literature assumes GTs dominate.
Approach
GNN+ wraps classic GNNs with six established components (edge features, batch normalization, dropout, residual connections, FFN blocks, and RWSE positional encoding), tuned per dataset, showing that simple GNNs with these components rival SOTA graph transformers.
Key results
- GatedGCN+ ranks top-three on all 14 datasets and first on 8, per the paper's summary; e.g., GatedGCN+ reaches 98.712% on MNIST and 77.218% on CIFAR10.
- GCN+ reduces ZINC MAE from 0.367 to 0.076 and raises PATTERN accuracy from 71.892% to 87.021%.
- GCN+ runs several times faster per epoch than GraphGPS on many datasets (e.g., ZINC 7s vs 21s, ogbg-code2 476s vs 1919s).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add edge feature integration (ablation) | GNN+ · ogbg-ppa | accuracy: 0.7310 to 0.5083 | Helped |
| add normalization (ablation) | GNN+ · ogbg-ppa | accuracy | Helped |
| add dropout (ablation) | GatedGCN+ · CIFAR10 | accuracy: 8.8% relative decrease | Helped |
| add residual connections (ablation) | GNN+ · CLUSTER | accuracy | Helped |
| add FFN (ablation) | GIN+ · MNIST | accuracy: 88% relative accuracy drop | Helped |
| add positional encoding (ablation) | GNN+ · ZINC | MAE | Helped |
| use GNN+ rather than vanilla GatedGCN | GatedGCN · ogbg-ppa | accuracy: 0.7531 -> 0.8258 | Helped |
| use GNN+ rather than vanilla GatedGCN | GatedGCN · Peptides-func | Avg. Precision: 0.6765 -> 0.7006 | Helped |
Surprising findings
- Retrained baseline graph transformers under the same search space did not surpass their originally reported results, so original-paper numbers were used.
- Dropout rates around 0.1 to 0.2 are optimal for graph-level tasks, contrasting with higher rates typically needed for node-level tasks.
Limitations
- Findings are purely empirical with no theoretical explanation for the advantage of GNN+.
- Current benchmark datasets may not capture the complexity of real-world graph problems, so the relative advantage may shift.
- Hyperparameters were not exhaustively searched due to the large search space.
Experimental setup
Datasets: ZINC MNIST CIFAR10 PATTERN CLUSTER Peptides-func Peptides-struct PascalVOC-SP COCO-SP MalNet-Tiny ogbg-molhiv ogbg-molpcba
Benchmarks: ZINC MNIST CIFAR10 PATTERN CLUSTER Peptides-func Peptides-struct PascalVOC-SP COCO-SP MalNet-Tiny ogbg-molhiv ogbg-molpcba
Metrics: MAE accuracy average precision F1 score AUROC time per epoch
Baselines: GraphGPS GRIT GEAET GRED TIGT Exphormer Graph-Mamba GSSC Graphormer SAN
Builds on: GraphGPS GatedGCN GIN GCN DeepGCNs
Closest papers by meaning
- Classic GNNs are Strong Baselines: Reassessing GNNs for Node Classification (NeurIPS 2024 · similarity 0.96)
- Graph Neural Networks Inspired by Classical Iterative Algorithms (ICML 2021 · similarity 0.92)
- Graph Neural Networks are Inherently Good Generalizers: Insights by Bridging GNNs and MLPs (ICLR 2023 · similarity 0.92)
- Making Classic GNNs Strong Baselines Across Varying Homophily: A Smoothness–Generalization Perspective (NeurIPS 2025 · similarity 0.90)
- Graph Neural Networks Are More Than Filters: Revisiting and Benchmarking from A Spectral Perspective (ICLR 2025 · similarity 0.90)
- A Scalable and Effective Alternative to Graph Transformers (AAAI 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.