The guide
Molecules, social networks, road maps and citation graphs do not fit in a
grid or a sequence. Graph neural networks work on them directly by
passing messages: each node collects the features of its neighbours,
combines them with its own, and updates. Stack k layers, and each node
sees k hops away.
The layers differ in how they combine messages. GCN takes a fixed, degree-normalised average. GraphSAGE keeps the node itself apart from its neighbours. GAT learns how much attention to give each neighbour. Every one of them must give the same answer however the nodes are numbered.
PageRank is here as the classical ancestor: a fixed message-passing rule, run to convergence, that ranks every node in a graph.