Editing Graph Neural Networks (section)

== <span style="color: #FFFFFF;">Evaluating</span> ==
Expert evaluation of GNNs requires care about what the performance numbers actually mean:

'''Transductive vs. inductive evaluation''': Transductive GNNs (GCN) train and test on the same graph with different node masks. Inductive GNNs (GraphSAGE) generalize to entirely unseen graphs or nodes. These are different capabilities — inductive evaluation is more practically useful and harder to fake.

'''OGB (Open Graph Benchmark)''': The standard benchmark suite for GNNs. Provides large, realistic graphs with standardized train/val/test splits and a public leaderboard. Much more rigorous than small academic datasets like Cora/Citeseer.

'''Link prediction pitfalls''': Naively splitting edges can cause data leakage (two nodes that appear together in training shouldn't appear as positive test pairs without careful subgraph isolation). The SEAL paper identified this issue; OGB provides proper splits.

Expert practitioners verify that their GNN is learning graph structure rather than node features alone — by ablating edge features, shuffling node labels, and measuring the performance delta. If removing edges doesn't hurt performance, the model isn't actually using graph structure.
</div>

<div style="background-color: #2F4F4F; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">