• the PR does not contain all changes mentioned in the preamble. Still WiP?

Yes, still work in progress.

• There is a similar tag to capability:variable_identification which already exists, capability:feature_importance, already - formally, the root cause or variable identification function is a 0/1 feature importance. We could therefore consider using that?

The nice thing about having a separate tag is that you can specify exactly what it means for detectors: That the output of predict will have an additional column named icolumns and that transform will have the same number of columns as the input, with a particular naming format. If that's not so important, it's nice to reuse the feature_importance tag I guess.

The nice thing about having a separate tag is that you can specify exactly what it means for detectors

You can still do this if you use the same tag - by tests or docstring, no?

But I do agree that it actually means something different for detectors, since we have feature importances per time index. A separate tag makes sense (I am now increasingly convinced of that) - although I wonder whether it can be shortened.

But I do agree that it actually means something different for detectors, since we have feature importances per time index. A separate tag makes sense (I am now increasingly convinced of that) - although I wonder whether it can be shortened.

• capability:subset_identification?
• capability:affected_components? (I've used this before.)

not much shorter...

capability:variable_attribution might be clearer than all the options?

Sorry, I forgot about following up this.

I think all our options here are pretty long, so therefore I'd go with the most self-explanatory name. I still think this is capability:variable_identification. If I was going to explain in a sentence what this capability means for a detector, I would write something like: "The detector identifies which variables are affected by the change/anomaly", rather than "The detector attributes the change/anomaly to a set of variables." If I have to use the "identify" term to explain the "attribution" term in the name, then I'd rather go with "identify" in the name.

What about capability:identify_variables?

What about capability:identify_variables?

I think the noun variant, i.e., capability:variable_identification, is slightly clearer.

Alright. Then I think this PR is ready.