Contribution
A controlled re-evaluation of popular GNNs and four new node-classification datasets. A benchmark framework came with it.
Key insight
A fixed test split plus model-specific training recipes measures overfitting to that split and the quality of the recipe. It does not measure the architecture.
Key results
- Average rank: GCN 2.3, MoNet 2.7, GraphSAGE-mean 2.7, GAT 3.6. The four baselines rank 7.4–8.8
- Planetoid split: GAT wins Cora and Citeseer. Another split of the same sizes: GCN wins both
- GAT collapses below 40% in 138 of 2,000 runs on Amazon Photo
Limitations
- Hyperparameters were selected on Cora and Citeseer only, then reused
- Two-layer models, transductive node classification, with small label budgets
- A workshop paper, and newer architectures are not covered