CaptumExplainer in Explqin do not support ShapleyValueSamplin ?

[arthurserres]

Hello, I am using Captumexplainer in order to explain prediction of my GNN. To be more precise, the configuration that I use is

explainer = Explainer( model=model, explanation_type='model', algorithm=CaptumExplainer('ShapleyValueSampling'), node_mask_type='attributes', edge_mask_type='object', model_config=model_config, threshold_config=dict(threshold_type = 'topk', value=topk) )

For some reason, computation of the explanation object is never ending although the graph is not that big... Is there a reason why I an stuck with this ? Maybe it is computationally hard, or simply this configuration does not make sense (not supported)?

Thanks in advance for the clarification :)

[rachitk]

ShapleyValueSampling is very slow, and on larger graphs or larger models, it can take a lot of time. Generally, this should work, but in practice it can simply just be time-consuming.

You can try other methods such as IntegratedGradients if the Shapley option is computationally intractable.

[arthurserres]

Hi !

Thanks a lot for this super fast answer, this is really helpful! Indeed, IntegratedGradients is quite fast but I do not manage to understand how the contributions are computed for the edges. I made a post about this here..

Very good to know that Shapley values approach should work,I'll just have to be patient.

Thanks :)