A Question about GCN Implementation in the examples

[shuoyinn]

Hello, your work is such a great contribution to the community and I really appreciate it since it is of much benefit to my study.

But for me there is a tiny question: I am confused about your implementation of GCN, why would you set a root_emb as a learnable embedding or parameter within a GCN Conv? I saw it is used to add to a linearly projected x before a relu followed by degree normalization. Or, I want to know what's the function and motivation of the variable root_emb. Is it the first time to use this variable in literature or you just follow someone's work?

I thought it was related to virtual node method, but virtual node was already implemented in GNN_node_Virtualnode, so I still don't know what's the meaning of root_emb here.

[weihua916]

Hi! Glad to hear you find OGB useful!

Note that the original GCN cannot incorporate edge feature information. Therefore, we extend the original GCN by making the message norm.view(-1, 1) * F.relu(x_j + edge_attr) (see here) instead of norm.view(-1, 1) * x_j. Then, you can think of self.root_emb as edge_attr analogue of the self-edge.

[shuoyinn]

@weihua916 Thanks for your such an prompt reply. Now I see maybe you want an analogue of x_j + edge_attr for self-loop (since self-loop is a default setting for a GCNConv), so you add x_i+edge_attr extra after aggregation of many x_j + edge_attr to realize it, where edge_attr for x_i's self-loop is not in raw data.edge_attr, so you just make it be a learnable or trainable. It helps a lot. Thanks, again =v=

[weihua916]

Exactly, that's the intention. You are welcome.