Inductive Representation Learning on Large Graphs

This paper presents GraphSAGE, a general inductive framework for generating low-dimensional node embeddings in large graphs. Whereas most prior approaches are inherently transductive — requiring all nodes to be present during embedding training and not naturally generalizing to unseen nodes — GraphSAGE leverages node feature information such as text attributes and, instead of training an individual embedding per node, learns a function that generates embeddings by sampling and aggregating features from a node's local neighborhood.

GraphSAGE outperforms strong baselines on three inductive node-classification benchmarks: it classifies the category of unseen nodes in evolving information graphs based on citation and Reddit post data, and it generalizes to completely unseen graphs on a multi-graph dataset of protein-protein interactions. This inductive capability matters because node embeddings had proved extremely useful in prediction tasks from content recommendation to identifying protein functions, and GraphSAGE extends them to previously unseen nodes and graphs.