Research map / Graph learning / GNN architectures
Classic GNNs are Strong Baselines: Reassessing GNNs for Node Classification
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 GCN, GAT, and GraphSAGE with careful hyperparameter tuning (normalization, dropout, residual connections, depth) on 18 node classification datasets, finding they match or beat recent graph transformers.
Problem
Whether graph transformers' reported superiority over message-passing GNNs for node classification is real, or an artifact of under-tuned GNN baselines.
Approach
Apply a tuned search space (normalization, dropout, residual connections, depth up to 10 or more layers) to classic GCN, GAT, and GraphSAGE, retrain graph transformer baselines under the same protocol, and run ablations of each hyperparameter across homophilous, heterophilous, and large graphs.
Key results
- Tuned GCN* reaches 85.10% on Cora and 46.29% on Chameleon (first place), and tuned GAT* reaches 85.01% on ogbn-proteins versus 72.02% previously reported for GAT.
- Classic GNNs rank top on 17 of 18 datasets, per the paper's claim.
- Ablations show normalization matters most on large graphs, dropout is consistently useful, and residual connections give the largest gains on heterophilous graphs (e.g., GCN* drops 16.43% on Roman-Empire without them).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Tune classic GCN hyperparameters | GCN · homophilous graphs | accuracy: 3.50↑ on Cora; 1.54↑ on CiteSeer; 2.32↑ on PubMed | Helped |
| Tune classic GNN hyperparameters | GCN, GraphSAGE, and GAT · heterophilous graphs | accuracy: 6.34↑ on Squirrel; 17.58↑ on Roman-Empire | Helped |
| Add normalization (ablation) | classic GNNs · large-scale graphs | ROC-AUC: 82.21 -> 77.42; 85.01 -> 80.32 | Helped |
| Add normalization (ablation) | classic GNNs · small homophilous graphs | accuracy | No clear effect |
| Add dropout (ablation) | classic GNNs · node classification datasets | accuracy: 79.72 -> 77.02 on PubMed; 91.27 -> 85.10 on Roman-Empire | Helped |
| Add residual connections (ablation) | classic GNNs · heterophilous graphs | accuracy: 91.27 -> 74.84 on Roman-Empire (GCN*); 90.63 -> 85.15 on Roman-Empire (GAT*) | Helped |
| Add residual connections (ablation) | classic GNNs · homophilous graphs | accuracy: 96.10 -> 94.43 on Photo; 96.17 -> 94.71 on CS | Mixed |
| Use deeper networks (ablation) | GCN* and GraphSAGE* · heterophilous graphs | accuracy | Helped |
| Increase depth beyond 10 layers (ablation) | GCN* and GraphSAGE* · Roman-Empire and Minesweeper | accuracy: 90.68 ± 0.44; 90.74 ± 0.38; 90.43 ± 0.52 | No clear effect |
| Use Jumping Knowledge aggregation (ablation) | classic GNNs · node classification datasets | accuracy | No clear effect |
Surprising findings
- Previously reported GNN baselines were substantially under-tuned; tuned GCN improves Roman-Empire by 17.58% over the originally reported result.
- Deeper GNNs (up to 10 layers, and 12-20 on some heterophilous graphs) help on heterophilous graphs, while the 56/112-layer DeepGCN/DeeperGCN configurations are unnecessary.
Limitations
- Only node classification is studied; graph-level and link prediction tasks are left out.
- Hyperparameter search for GTs and GNNs is constrained to a predefined space, and JK tuning was not exhaustive in earlier experiments.
Experimental setup
Datasets: Cora CiteSeer PubMed Computer Photo CS Physics WikiCS Squirrel Chameleon Roman-Empire Amazon-Ratings
Benchmarks: Cora CiteSeer PubMed Computer Photo CS Physics WikiCS Squirrel Chameleon Roman-Empire Amazon-Ratings
Metrics: accuracy ROC-AUC
Baselines: GraphGPS Exphormer NAGphormer GOAT NodeFormer SGFormer Polynormer H2GCN
Builds on: GCN GAT GraphSAGE Polynormer DeepGCN
Closest papers by meaning
- Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 2025 · similarity 0.96)
- Diffusing to the Top: Boost Graph Neural Networks with Minimal Hyperparameter Tuning (ICLR 2025 · similarity 0.91)
- Graph Neural Networks Are More Than Filters: Revisiting and Benchmarking from A Spectral Perspective (ICLR 2025 · similarity 0.90)
- Making Classic GNNs Strong Baselines Across Varying Homophily: A Smoothness–Generalization Perspective (NeurIPS 2025 · similarity 0.90)
- GC4NC: A Benchmark Framework for Graph Condensation on Node Classification with New Insights (NeurIPS 2025 · similarity 0.89)
- NLGT: Neighborhood-based and Label-enhanced Graph Transformer Framework for Node Classification (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.