Inductive Representation Learning on Large Graphs

Inside the paper

The problem

Many graph embedding methods learn a separate representation for each node in a fixed graph. That makes it awkward to represent newly arriving nodes or nodes in an entirely unseen graph without further optimization.

The idea

GraphSAGE learns functions that build a node representation from its features and a sampled neighbourhood. The same learned functions can then be applied to nodes the model did not encounter during training.

How it works

• Sample a bounded set of neighbours at each layer so that the computation for a node does not expand to its entire neighbourhood.
• Aggregate the sampled neighbours’ current representations, combine that summary with the node’s own representation, and transform the result with learned parameters.
• Repeat the procedure across layers to incorporate information from progressively farther away; the paper compares mean, pooling and other aggregation functions.
• Train either with a graph-based unsupervised objective or directly with labels, then use the learned aggregation procedure to embed unseen nodes.

What it showed

The authors evaluate on citation, Reddit and protein interaction data, including protein graphs held out during training. Their reported results show improvements over the feature and embedding baselines in these settings, with performance varying by aggregation and training method.

Read it for

Figure 1 and Section 3 explain the sampling and aggregation procedure; Table 1 and Section 4 put its inductive claims to the test. Skim the detailed dataset descriptions first if you want the architecture before the experimental setup.

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