Theoretical analysis of the representation power of a 1-layer GCN for non-i.i.d. semi-supervised node-level tasks.
I can't tell for now, but it's interesting for sure...