[RFC] Voting classifier flatten transform (#7230) #7794
Conversation
|
Also, I'm not sure whether regression test is needed. Because added functionality is pretty simple. |
|
I'd be tempted to deprecate the current behaviour. |
|
tests are always needed! |
|
Do you see harm in deprecating entirely? I'd rather that. The current |
|
I'm happy with deprecating entirely. That needs two steps, though: introducing the parameter and removing it again. Or what did you want to do? |
|
Also, it might be helpful to special case for binary classification and only retain probabilities of one of the classes, say the positive one? That would result in smaller and more interpretable downstream models. |
|
bonus points for implementing |
|
|
||
|
|
||
| def test_transform(): | ||
| """Check trqansform method of VotingClassifier on toy dataset.""" |
| flatten_transform=True).fit(X, y) | ||
|
|
||
| assert_array_equal(eclf1.transform(X).shape, (3, 4, 2)) | ||
| assert_array_equal(eclf2.transform(X).shape, (4, 6)) |
There was a problem hiding this comment.
can you check more than shapes with an assert_equal on the values after a proper reshape?
| flatten_transform : bool, optional (default=False) | ||
| Affects shape of transform output only when voting='soft' | ||
| If voting='soft' and flatten_transform=True, transform method returns | ||
| matrix with shape [n_samples, n_classifiers * n_classes] instead of |
There was a problem hiding this comment.
[n_samples, n_classifiers * n_classes]
->
(n_samples, n_classifiers * n_classes)
shapes are tuples
| Affects shape of transform output only when voting='soft' | ||
| If voting='soft' and flatten_transform=True, transform method returns | ||
| matrix with shape [n_samples, n_classifiers * n_classes] instead of | ||
| [n_classifiers, n_samples, n_classes]. |
| @@ -238,16 +247,25 @@ def transform(self, X): | |||
|
|
|||
| Returns | |||
| ------- | |||
| If `voting='soft'`: | |||
| If `voting='soft'` and `flatten_transform=False`: | |||
| array-like = [n_classifiers, n_samples, n_classes] | |||
| - Added ``flatten_transform`` parameter to :class:`ensemble.VotingClassifier` | ||
| to change output shape of `transform` method to 2 dimensional. | ||
| (`#7794 <https://github.com/scikit-learn/scikit-learn/pull/7794>`_) | ||
| by `Ibraim Ganiev`_. |
There was a problem hiding this comment.
should be written as
:issue:7794 by Ibraim Ganiev_.
|
How does this relate to our concurrent ambitions towards stacking (e.g. #8960) |
|
You'll also need to rebase on master, to get the tests running. |
Reference Issue
Fixes #7230
What does this implement/fix? Explain your changes.
It adds
flatten_transformparameter toVotingClassifier, which changes shape of transform method's output to[n_samples, n_classifiers * n_classes]instead of[n_classifiers, n_samples, n_classes],With this parameter turned on you can use
VotingClassifieras a transformer, and feed its output to other estimators/transformers inPipeline.Any other comments?
None, make suggestions. I Summon @amueller into this PR :)