Edge weights in GIN

[mdanb]

How can I use edge weights with GINConv? I saw that GCN conv has an optional attribute which allows you to use graph weights. Is it not possible to do the same with GINConv?

[reshinthadithyan]

I guess, you can use torch_geometric.nn.conv.GINEConv.

[mdanb]

Thanks didn't notice that! haha

[simon-forb]

Note that the suggested model would not use the edge weights as weights but rather as attributes. What this means is that the edge attributes are added to the node attributes in the message passing step. Naturally one would expect that the edge weights would serve as a scaling factor, e.g., if an edge has weight $1/2$, then the node attributes from the connected node would by scaled by $1/2$.

For instance see the implementation in DGL: https://docs.dgl.ai/en/0.8.x/generated/dgl.nn.pytorch.conv.GINConv.html

[rusty1s]

It does not make much sense to add edge_weight support to GINConv since this conflicts with the formulation of GIN in the first place (injectivity is no longer preserved).