Add specific tags for Search and Ensemble classes (and maybe Transformers)

I am using sklearn classes extensively for a project. I am wrapping all the classes to extend the basic sklearn functionalities using class decorators.

I find very useful to have all the information extracted thanks to _get_tags method and also to the _estimator_type attribute which lets me customize the classes depending on some criteria.

I would have thought that this should be the case also for Ensemble classes like BaggingRegressor and Search classes like GridSearchCV.

At the moment I am using a trick to understand if a particular class is a Search or Ensemble but this might not work in the future.

For example for Search classes I check if the instance has the estimator attribute which worked so far even with other libraries which implement the sklearn API like for example scikit-optimize and its BayesianSearchCV.

But I believe it would be useful for developers to add this information to the sklearn API maybe using an additional tag in the base classes for Search and Ensemble i.e. BaseSearchCV and BaseEnsemble.

Mayne it can also help to have the tag in the Transformer class to understand if it’s a transformer such that one doesn’t have to rely on the “unwritten” rule of checking if an instance it has the predict method

[gioxc88]

I don’t think tags will solve your issue: third party developers would still need to use / set those tags in some way, the same way they should / would use the current scikit-learn Mixins / base classes.

In other words, just because the tag exists in sklearn doesn’t mean that HyperbandSearchCV would have its tag set to the proper value.

Thank you for the answer, I am well aware of this. My example in fact had the objective of emphasising what you said: it should be up to developers to use tags on inherit from the the relevant class.

My point was to deprecate inheriting from MetaEstimatorMixin and adding tags to BaseSearchCV and BaseEnsemble.

[NicolasHug]

Since the original comment was addressed I’ll close the issue