minor: Causal fidelity metrics: Harmonization between timeseries and image - detailed evaluate

[AntoninPoche]

minor: Causal fidelity metrics: Harmonization between timeseries and image - detailed evaluate

This pull request is divided in 3 contributions :

• Harmonization between:
  • CausalFidelity and CausalFidelityTS edcb5c4
  • Adapt testing to harmonization 03d3849
• Detailed evaluate for:
  • CausalFidelity e059af8
  • CausalFidelityTS 4351630
• Remarks in the docs for some metrics:
  • Insertion and Deletion 4e177cc
  • InsertionTS and DeletionTS af217c7

and a correction for the new version of pylint to work: 22aebd4 (we should use f-string).

Harmonization

The previously developed causal fidelity metrics for timeseries had several differences in their behavior compared to the initial Insertion and Deletion, thus an harmonization was necessary.

Detailed evaluate

Furthermore, through discussion, the necessity of accessing all values computed for insertion and deletion was necessary. The idea thus was to add a method (detailed_evaluate) with returns a dictionnary of such values. This method will then be called by the initial evaluate method that will return the auc.

The output of the detailed evaluate method can be used to draw score evolution curves, where the score depend on the number of features perturbed.

Interpretation and remarks in the documentation

There are remarks about:

• The score interpretation
• How max_percentage_perturbed can influence the score
• The fact that evaluating Insertion and Deletion on the absolute value of explanations may not provide relevant scores.

[fel-thomas]

You're trying to re-commit commits from last month, can you clean your PR with just the new commits pls ? ;)

[lucashervier]

Most of my remarks question the relevance of some change to get more complete metrics. The idea is good but there is a few change to make I believe or to discuss at least. Otherwise, well done 👍 🎉

[fel-thomas]

Cool PR, LGTM ! ;)

[lucashervier]

Thank you for the modification!