giskard
giskard copied to clipboard
Giskard-AI
Scan: Add a robustness detector to the scan that perturbs categorial values
🚀 Feature Request
Add a robustness detector to the scan that perturbs categorial values.
The scan should be able to a set of issues that capture the perturbations needed on a single categorial feature to:
(a) change the predicted label (classification) (b) change the prediction by an amount that exceeds a certain threshold (regression)
🔈 Motivation
Currently the scan does not have any categorial perturbation.
@kevinmessiaen I let you guide there, this seems easy to add, and a great idea of contribution!
Hello @ChatBear
Yes this is still an active issue, I can assign you to it. We would be grateful to have your contribution, let me know if you have question about this.
Thanks, i'll try to contribute, i'll need a bit of time to understand the repo, after that i'll try to post PR
Hello, i have few questions about the issue.
What kind of pertubations do you except ? I was thinking of change the feature column with a probability of 0.1 (chosen arbitrary).
And do i need the create another detector from scratch, or i can use a detector from BaseTextPerturbationDetector ?
And i tried to create a branch, and i can't push in my own branch (i forked the repo but i am having trouble to create the pull request, i am kinda of new in open source so i apologize in advance if this question is inappropriate).
Hello,
The perturbation should be on categorical feature. It should only perturb on column of the dataset, the goal is to ensure that the model isn't too sensitive to noise. In this case the probability is not necessary since we want to test that the result isn't impacted when the value change. (it makes sense in text where we have typo rate for example).
Example is having a breed category with values potential values: ['Labrador', 'Husky', 'Beagle', ...]. The idea is to switch all Labrador` value to any other breed and so on.
It won't work to reuse BaseTextPerturbationDetector since it cast column as str but we can have numerical categories for example. But you can inspire from it.