When to Use GNNs & Evaluation

Honest decision guide - GNN vs tabular ML vs LLM, metrics, baselines, and common evaluation mistakes.

1. Decision Framework

Ask these questions in order:

  1. Is there relational structure? If rows are independent (customer churn with no links), use XGBoost/MLP first.
  2. Do labels depend on neighbors? If yes (homophily or structured heterophily), GNN is a strong candidate.
  3. Do you have node features? Structure-only → random walk embeddings or label propagation may suffice.
  4. Graph size? Small → full-batch GCN. Large → GraphSAGE + neighbor sampling.
  5. Dynamic? Static → standard GNN. Streaming edges → TGN.
  6. Multiple types? Heterogeneous → R-GCN/HGT.
  7. Text on nodes? Consider LLM embeddings + GNN, or GraphRAG - see GNNs & LLMs.

2. When NOT to Use GNNs

Rule of thumb: Always run an MLP on node features and label propagation on edges. Your GNN must beat both to justify complexity.

3. GNN vs MLP vs LLM

ApproachWins whenLoses when
MLP / XGBoostFeatures carry most signal, no neighbor effectStructure is key (citations, molecules)
GNNLocal structure predicts labels, inductive generalization on graphsLong text reasoning, no graph
LLM aloneText-heavy, few-shot reasoning, no reliable graphLarge structured graphs, combinatorial structure
LLM + GNNText-attributed graphs, KG + language queriesSimple homophilous node cls with bag-of-words features

4. Metrics

TaskMetrics
Node classificationAccuracy, F1 (macro for imbalanced), AUC-ROC
Link predictionAUC-ROC, AP, MRR, Hits@10, Hits@50
Graph classificationAccuracy, F1, AUC
Graph regressionMAE, RMSE, $R^2$

Hits@K: fraction of true edges where the correct target ranks in top $K$ predictions.

5. Baselines You Must Beat

6. OGB and Fair Comparison

Open Graph Benchmark provides fixed splits, evaluators, and leaderboards. Use their splits - do not invent your own when comparing to published numbers.

Report: mean ± std over multiple seeds (at least 3, preferably 10).

7. Ablation Studies

Test what actually matters: