Non-symmetric normalisation for directed graphs GNN

[seven-down]

I have a set of directed graphs and have read that non-symmetric normalisation (D^-1A) of the adjacency matrix is preferable when building a GNN. There doesn't appear to be any transformers available for this normalisation and the GCNConv layer normalizes symmetrically (D^-1/2AD^-1/2).

Any advice or resources would be greatly appreciated :)

[EdisonLeeeee]

I think part of the code in get_laplacian is what you are looking for.

[rusty1s]

In addition, D^-1 A can be simply achieved via aggr="mean" in the message passing layers, so there is no good reason to pre-compute this in the first place.

[seven-down]

Many thanks for the reply! Just to clarify, using aggr="mean" in the message passing layers does this need to be applied to each layer or only the first convolutional layer?

[EdisonLeeeee]

In your case it should be applied to each layer, I think.

[seven-down]

That makes sense, thank you very much for the help :)