Giters

fel-thomas / Sobol-Attribution-Method

👋 Code for the paper: "Look at the Variance! Efficient Black-box Explanations with Sobol-based Sensitivity Analysis" (NeurIPS 2021)

Home Page:https://github.com/deel-ai/xplique

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

about signed total Sobol index and first-order Sobol index

abcde-1447 opened this issue · comments

commented

Hi, it's me again. I have been using your method for interpreting a few of my models, so far it works great!

I noticed that you mentioned signed total Sobol index and first-order Sobol index, I am just wondering whether they are implemented somewhere in the code?

Thanks!

commented

Hi, ;)
Thanks !!! The First-order indices are not implemented here to encourage the use of total indices (the first order indices get much less interesting results). Concerning the signed estimator, you just need to multiply the total order indices with the results of the Occlusion method (take the Sign(.) of the results) -- available with the Xplique library Occlusion documentation.

e.g : Sti_delta(x) = Sign(Occlusion(x)) * Sti(x)

Nevertheless we have ideas to make the signed estimator more intelligent and a more elegant solution should come when we merge the Sobol method into the Xplique library.

commented

@fel-thomas Got it, thanks for the explanation!