Semi-Supervised Classification with Graph Convolutional Networks

Inside the paper

The problem

When only a few nodes in a graph have labels, their links and features may still help classify the rest. Existing graph methods could be expensive or require separate stages for learning representations and making predictions.

The idea

The authors build a neural network whose layers combine each node’s features with those of its neighbours. They derive a simple propagation rule from an approximation to spectral graph convolutions and train it on the labelled nodes.

How it works

• Add a self-link at each node and normalize the adjacency matrix so that information from neighbours can be combined with a node’s own features.
• Each layer applies this graph propagation to node features, then a learned transformation and a nonlinear activation.
• A final classifier produces node label probabilities, and the loss uses only the nodes whose labels are supplied during training.
• Stacking layers allows information to travel beyond immediate neighbours, while the sparse graph structure keeps propagation tied to the number of edges.

What it showed

On citation networks and a knowledge graph, the reported classification accuracies exceed the compared baselines on the paper’s standard splits. The authors also compare propagation choices and training times, so the results distinguish the model from its alternatives in more than one way.

Read it for

Read Section 2 for the derivation and Figure 1 for the network, then consult Section 6 and Table 2 for the classification comparison. The spectral approximation details can be skimmed initially and revisited after the propagation rule is familiar.

Written in our own words from the paper itself. A draft the editors have not yet reviewed.