Reference Issues/PRs

• Closes prefit option missing for voting? #12297

What does this implement/fix? Explain your changes.

This feature enables the user to pass in prefit estimators to both VotingClassifier and VotingRegressor by specifying prefit=True.

The VotingX is still required to have fit() called on it to do validation of the prefit estimators but it does not perform model fitting during this period.

Along with this, I've provided some basic first testing to ensure that the prefit estimators work equally if fit outside vs inside the Voter and validation checks during the call to VotingX.fit() to ensure that that prefit estimators are aligned in terms of the data that can flow through them.

Any other comments?

This is mainly a draft PR to backup my comment left on issue #12297 with an initial implementation. There are likely a lot of missing features and issues that could arise.

Below is a list of issues that need to be resolved for this feature to integrate properly or issues that need to be discussed.

Issues

• When calling VotingX(estimators, prefit=True).fit(X, y), should the Voter validate that the estimators are aligned with eachother, or with the passed in X, y, in terms of attributes like n_features_in_ and n_feature_names_?
• What should be the defined clone behavior of this estimator, in relation to this comment by @amueller, linking to API Freezing estimators #8370
• What tags should be attached to the prefit version of a Voter with respect to the suite of estimator tests that exist for scikit-learn integrations.

As a meta-level question, should there exist a separate PrefitVotingX to alleviate some of this issues and simplify internals, at the expense of introducing a new estimator class?

TODOs if accepted

• Add a ..versionadded:: to the docstrings
• Add a changelog entry