How are contributions computed with the explain class

[arthurserres]

Hello all,

I try to understand the way the contributions are computed by the explain library.

I use Captum explainer and in particular Integrated Gradient method. I read the paper explaining the integrated gradient method so the method by itself is clear. However I looked at the source code as well at the documentation but I do not manage to understand how each node feature abd edge contribution are computed, may someone help on that ?

Thanks a lot :)

[rusty1s]

For integration with GNNs, we wrap the utilized GNN into a parent module, which expects two inputs: The node feature matrix, and a soft edge mask which is multiplied within message passing to every message. These are the two inputs that Captum will generate attributions for. Hope this answers your question.

[arthurserres]

Thanks for the clarification, I think it helps me quite alot but I oppened another discussion in order to precise my question: here
I would really appreciate if you manage to give an answer :)