Thanks for the links, Bernardo. Indeed, there is some variability in the terminology as the field is rather fragmented. GCN is a particular architecture due to Kipf&Welling built of two operations: node-wise linear feature transformations and graph diffusion with a constant operator. This is the simplest analogy of classical convolutions on graphs, and has deep roots in graph signal processing (one can e.g. think of graph adjacency matrix as an analogy of the shift operator)