When Design Rules Break: Benchmark Composition Determines Whether Label Informativeness Predicts GNN Aggregator Choice
Researchers investigated how graph neural network (GNN) design rules perform across different benchmark datasets. They found that the composition of these benchmarks significantly impacts whether label informativeness predicts the choice of GNN aggregators. Specifically, dense friendship networks like Facebook-100 showed a preference for sum aggregation even with low label informativeness, a behavior not replicated by standard models or degree-based metrics alone. The study suggests that benchmark diversity, rather than just numerical insufficiency, determines the apparent generalization of design rules. AI
IMPACT Highlights the importance of diverse datasets for evaluating and developing graph neural network models.