GCNII

Every practitioner knew the symptom: stack more than two or three GCN layers and accuracy collapses, because repeated smoothing makes all node embeddings converge. GCNII diagnosed the cause and fixed it with two borrowed tricks. An initial residual connection carries the input features forward to every layer, so deep representations can always fall back on the raw signal; an identity mapping added to each weight matrix keeps the layer close to the identity function at initialization, so depth stops destroying information. The authors also provide a K-layer analysis showing GCNII provably expresses a K-order polynomial spectral filter.\n\nThe ICML 2020 paper reports 85.5% on Cora with 64 layers, 73.4% on Citeseer with 32 layers and 80.2% on Pubmed with 16 layers, beating GCN, GAT, APPNP and JKNet on all three. The contrast with the plain GCN is stark: where the vanilla model falls to near-random accuracy at 64 layers, GCNII holds steady. GCNII reopened depth as a design dimension for GNNs and its initial-residual idea shows up in many later architectures.

Paper-reported: 85.5% Cora (64 layers), 73.4% Citeseer (32 layers), 80.2% Pubmed (16 layers), beating GCN, GAT, APPNP and JKNet on all three (Chen et al., ICML 2020).