#gat

Question 1: Single choice

How does a GAT layer weight the neighbours of a node differently from a GCN layer?

GAT learns them from node features by attention; GCN fixes them from degrees1
GAT attends to every node of the graph, while GCN uses only direct neighbours2
GAT weights neighbours with eigenvectors of the graph Laplacian, GCN with degrees3
GAT gives every neighbour the weight $1/|\mathcal N(i)|$ , while GCN learns its weights4