When to Use GNNs
When should you use a GNN?
When data has meaningful relational structure, labels/features of a node depend on its neighbors, you have node features and edges, and simpler baselines (MLP, label propagation) are beaten by structure-aware models.
When should you NOT use a GNN?
i.i.d. tabular data with no real relationships, invented k-NN graphs with no semantic meaning, tiny graphs, when MLP on features already wins, strict latency on 100M-node graphs without sampling infrastructure.
GNN vs XGBoost on tabular data with a graph?
If graph structure adds signal beyond features (citation topic, social influence), GNN wins. If features are sufficient and edges are weak, XGBoost often wins with less complexity. Always ablate: GNN vs MLP vs XGBoost.
What metrics for link prediction?
AUC-ROC, Average Precision (AP), MRR (mean reciprocal rank), Hits@K (fraction where true link ranks in top K). Report on filtered setting (hide other true edges when ranking).
What is the OGB benchmark?
Open Graph Benchmark - standardized datasets, splits, and evaluators for fair comparison. Use their splits and report mean ± std over multiple seeds. Leaderboard at ogb.stanford.edu.
What ablations should every GNN paper include?
MLP baseline (no graph), number of layers, aggregator type, with/without edge features, with/without positional encodings. Honest comparison builds trust.