Graph Isomorphism Network (GIN)

GIN answered a theoretical question the field had been avoiding: how powerful can a message-passing GNN actually be? Xu and colleagues proved that a GNN can be at most as discriminative as the Weisfeiler-Lehman graph isomorphism test, and that reaching that ceiling requires injective neighborhood aggregation. Mean and max pooling fail the injectivity condition, so GCN and GraphSAGE provably cannot distinguish certain simple graph structures. GIN's fix is deliberately simple: sum the neighbor features and pass them through a multi-layer perceptron, which the universal approximation theorem makes injective. A learnable epsilon weights the central node.\n\nThe ICLR 2019 paper shows GIN fitting training data almost perfectly where weaker variants underfit, and reaching state of the art on multiple TU graph-classification benchmarks. It later became the default OGB molecular baseline: GIN with a virtual node scores 77.07% test ROC-AUC on ogbg-molhiv. GIN reframed how the field thinks about architecture design, and its WL-equivalence argument is now standard textbook material.

Paper-reported: state of the art on multiple TU graph-classification benchmarks (Xu et al., ICLR 2019). Independent OGB baselines: GIN + virtual node reaches 77.07% test ROC-AUC on ogbg-molhiv (Hu et al., 2020).