GNNExplainer inconsistency with the paper?

[DimGorr]

🚀 The feature, motivation and pitch

Hi! I have a question about GNNExplainer feature selection part. In the article it's claimed that in order to solve the potential issue with importnat features whose values are zeros 1) Monte Carlo Estimate to sample feature subset and then you use 2) reparametrization are used

The problem is that I actually haven't found those parts in pytorch geometric implementation. In fact, I'm asking this to find out if pytorch geometric could mistakenly show that some feature is unimportant if it was equal to zero:)

Thank you in advance!

Alternatives

No response

Additional context

No response