Isn't there a chance we'll need ChangedBehaviorWarning in future? I don't remember what it was used for, maybe @amueller @jnothman @ogrisel @rth do?

My impression is that we eventually decided to use FutureWarning for all of them.

NonBLASDotWarning seems to have been used when numpy.dot had speed issues (version < 1.7). I think it has no utility anymore and can be removed safely.

You can remove this filter as well https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/utils/validation.py#L36

@jeremiedbb the filter is already removed in this PR, or at least I think it's removed. Don't you see it in the diff?

I specifically checked if you already did it and yet I missed it. The diff is too big :D

The ChangedBehaviorWarning was something we did after changing the behavior, so it's unlike future warning, but I also think it's not a good idea.

#17610 shows a use-case where ChangedBehaviorWarning would make sense. Basically, it would be linked to changing the default of a parameter for an experimental function/class. The reasoning of doing such was brought by @jnothman with who I agree:

I think forcing the user to always set as_frame is excessively conservative, and creates unnecessary as_frame= where the dataset would fail to load with as_frame=False (e.g. string columns).

I might think that we might need this warning at some point in HistGradientBoosting or IterativeImputer, for instance.

I'd say if we think we should give users warning about a change behavior in an experimental module, we're not thinking of it really as experimental and we should move it to stable, and deprecate old behavior as usual. I'd find it odd if we go down the path of warning users about change behavior in a module which is explicitly there for us to change things w/o a warning.