Summary

This chapter introduced the GraphSAGE framework and its two components – the neighbor sampling algorithm and three aggregation operators. Neighbor sampling is at the core of GraphSAGE’s ability to process large graphs in a short amount of time. It is also responsible for its inductive setting, which allows it to generalize predictions to unseen nodes and graphs. We tested a transductive situation on PubMed and an inductive one to perform a new task on the PPI dataset – multi-label classification. While not as accurate as a GCN or a GAT, GraphSAGE is a popular and efficient framework for processing massive amounts of data.

In Chapter 9, Defining Expressiveness for Graph Classification, we will try to define what makes a GNN powerful in terms of representation. We will introduce a famous graph algorithm called the Weisfeiler-Lehman isomorphism test. It will act as a benchmark to evaluate the theoretical performance of numerous GNN architectures, including...