torch_geometric.explain and edge features

[arthurserres]

Hello,
I am trying to explain my GNN using the explain experimental module. All the features that I use in my model are on the edges and not on the nodes. I thus wanted to consider the mask "object" for node and "common_attributes" which is the correct configuration If I goodly understood the use of masks. However I looked at the source code and understand that edge mask can only be None or "object", so if I understand right, edge attribute cannot be used to explain the outcomes, is that correct ?
Thanks in advance.

[rusty1s]

Yes, this is correct. While it is not impossible to support explanations of edge features, it makes the implementation a bit harder since we now need to separate the implementation of explainers based on the edge_mask property, i.e., edge_mask=object refers to explanation of edge_index, while edge_mask=common_attributes refers to explanation of edge_attr.

[arthurserres]

I see, thanks a lot for the update it's nice of you!
I saw that Captum can natively be used to explain GNN (I think it essentialy does what we can do with the captum explainer and the explain class, correct ? ).
Is there a chance I can easily modify this notebook or the source code of explain such that it also produces explaination using edge features ? Or is it a more complicated thing that requires high level coding skills ?

Thanks by advance :)

[rusty1s]

What we need to change is the to_captum_input function, which also respects edge_attr as an input feature and includes it into its attribution.

[arthurserres]

Thanks a lot, I will have a look :)