To be used sensibly in a Pipeline / parameter search, an estimator's transform method can't have parameters other than the data (though it may be possible in the future to support more than one data argument). Rather, it should be possible to set the parameter using the estimator's set_params method as used in a parameter search.

Hence additional arguments to transform should become object parameters (though whether they should be made unavailable as arguments directly to transform is up for debate).

Unfortunately, these happen to appear in mixin classes, meaning a lot of sub-classes' docstrings need altering.

Examples of additional arguments:

• scikit-learn/sklearn/cluster/_feature_agglomeration.py

  Line 22 in 71d00d9

  def transform(self, X, pooling_func=np.mean):

• https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/feature_selection/from_model.py#L20 should have selection_threshold parameter on object