@glemaitre, thank you for this answer and the reminder on the purpose of TargetTags.

Just to be absolutely clear, did you mean "does not use y and thus does not use the TargetTags"?

For our transformers that don't use y, would you recommend that we set TargetTags to None? Because our transformers have BaseEstimator as a base class, I assume we'd need to do something like:

class CustomTransformer(TransformerMixin, BaseEstimator):
    def __sklearn_tags__(self):
        tags = super().__sklearn_tags__()
        tags.target_tags = None
        # Other tag settings ...
        return tags

or just leave it to the TargetTags as defined in BaseEstimator?

Do you anticipate that transformers will eventually have estimator_type="transformer" and should we stick with estimator_type=None in our transformers until then?

In the near future, estimator_type we don't plan estimator_type="transformer" because there is case where there is actually several type possible as mentioned here: #30368 (review)

So I anticipate that having None for the moment is safe. We might extend the API later but it will not be with a breaking change.