Why need to add alpha_j and alpha_i in GATConv ?

[Nassiming]

Hi, I'm a rookie to PyG. I'm confused about a line in the GATConv source code:

# we simply need to sum them up to "emulate" concatenation:
alpha = alpha_j if alpha_i is None else alpha_j + alpha_i

I don't understand why we need to add alpha_j and alpha_i. And I would greatly appreciate it if someone could provide me with additional explanations

[rusty1s]

alpha_j refers to the attention scores of the source node, and alpha_i refers to the attention scores of the destination nodes, which gets combined via summation. This is identical to first concatenation source and destination node features, and then applying multiplying that by an attention vector. However, this scheme is more efficient since we avoid the expensive concatenation here.