GNN Aggregator Design Rules Don't Generalize Across Benchmarks

A new paper from Neha Sharma and Ritesh Sharma challenges widely held assumptions about graph neural network (GNN) aggregator selection. By evaluating sum, mean, and max aggregators across 24 node-classification datasets—including citation, heterophilic, LINKX, co-purchase, co-authorship, and Facebook-100 graphs—they demonstrate that design rules are far from universal. The commonly used metric of label informativeness predicts performance gaps between GIN-Sum and GIN-Mean reliably on legacy benchmarks (like Cora and Citeseer), but breaks down entirely when Facebook-100 graphs are included. In these dense friendship networks, near-zero label informativeness coexists with a strong preference for sum aggregation, yielding 7-10% improvements and up to 13% under extended training. Stochastic block model ablations that matched Facebook-100 degree scales failed to reproduce this effect, ruling out mean degree as the sole explanation.

Among several label-independent graph statistics, the spectral gap emerged as the only feature uniquely distinguishing Facebook-100 from other low-informativeness datasets, with the effect localized to one-hop neighborhoods and replicated across different GNN architectures. The study also finds that the PNA (Principal Neighbourhood Aggregation) layer can underperform the best single-aggregator GIN on standard citation benchmarks, contradicting previous claims about its superiority. These results imply that benchmark composition—not numerical insufficiency—determines whether GNN design rules appear to generalize. The authors argue that the Facebook-100 regime provides a concrete target for developing future adaptive aggregation methods, urging researchers to test across diverse graph families before drawing conclusions about optimal aggregation strategies.