In 2016 Thomas Kipf and Max Welling introduced the Graph Convolutional Network (GCN), a model that runs convolution-like operations directly on graph-structured data such as citation networks and social graphs. The paper, “Semi-Supervised Classification with Graph Convolutional Networks,” was submitted to arXiv on September 9, 2016. Its central idea is to derive a simple layer-wise propagation rule from a localized first-order approximation of spectral graph convolutions, which makes the model both efficient and easy to implement.
A GCN layer updates each node’s representation by mixing its own features with those of its neighbors, then passing the result through a learned weight matrix and a nonlinearity. Stacking a few such layers lets information flow across the graph so that a node’s final embedding reflects both its own attributes and the structure of its local neighborhood. The authors showed the approach scales linearly in the number of graph edges.
The model was designed for semi-supervised learning, where only a small fraction of nodes carry labels. Tested on citation networks and a knowledge graph dataset, the GCN outperformed existing methods by a clear margin while remaining computationally cheap. Because the formulation was so compact, it became one of the most widely cited and reused graph learning baselines.
For a business reader, the GCN is the practical entry point to learning on networks: customer graphs, supply chains, fraud rings, and molecular structures are all graphs, and this paper showed that a small, fast model could learn useful patterns from them with very little labeled data.