Graph Convolutional Network (GCN)

GCN is the paper that turned graph deep learning into a mainstream discipline. Kipf and Welling showed that a localized first-order approximation of spectral graph convolutions, combined with a simple renormalization trick on the adjacency matrix, produces a model that trains in linear time in the number of edges and learns representations fusing local graph structure with node features. The canonical two-layer architecture (16 hidden units, dropout, weight decay) became the default template for thousands of follow-up papers.\n\nThe ICLR 2017 paper reports 81.5% on Cora, 70.3% on Citeseer, 79.0% on Pubmed and 66.0% on the NELL knowledge graph, beating every baseline of its era by a clear margin. It is now one of the most-cited machine learning papers ever written, and its layer (GCNConv) ships in PyTorch Geometric and DGL. Later work showed its limits too: accuracy degrades fast past a few layers (over-smoothing), and fixed normalized-adjacency weights leave performance on the table versus learned attention.

Paper-reported: 81.5% Cora, 70.3% Citeseer, 79.0% Pubmed, 66.0% NELL (Kipf & Welling, ICLR 2017). Independently replicated many times; community reproductions in PyTorch Geometric land at ~81-82% on Cora.