On 02/02/2012 12:34 PM, Olivier Grisel wrote:
> 2012/2/2 Mathieu Blondel<[email protected]>:
>    
>> On Thu, Feb 2, 2012 at 8:15 PM, Olivier Grisel<[email protected]>  
>> wrote:
>>
>>      
>>> I wonder which representation is the nicest for the end user? It might
>>> be the case that keeping the unlabeled data as a separate variable
>>> might be more natural but that will probably impact pipeline-ability
>>> and cross-validation since X_unlabeld.shape[0] won't be the same as
>>> X_labeled.shape[0] and y_labeled.shaped[0].
>>>        
>> cross-validation will probably break any way as the unlabeled examples
>> cannot be used in the test set. This is also shows that we should
>> probably have a library-wide default encoding for unlabeled data (this
>> way, we will be able to make sure that all the unlabeled data goes to
>> the training set).
>>
>> Keeping the label propagation and semi-supervised NB PRs on hold
>> forever doesn't help. We should merge them and keep in mind that their
>> API is a work-in-progress.
>>      
> Alright: this should be made explicit in the whats_new.
>
>    
Larsmans pointed out that the -1 is already used in
One-Class SVM (and I think one other place).

I agree that there is a conflict between using -1 as unlabeled
and the use in binary classification.

But I also think this is not really part of the PR.

I would consider changing the unlabeled marker to -2,
but then after merging the two semi-supervises algos
and doing it in a separate PR.


------------------------------------------------------------------------------
Keep Your Developer Skills Current with LearnDevNow!
The most comprehensive online learning library for Microsoft developers
is just $99.99! Visual Studio, SharePoint, SQL - plus HTML5, CSS3, MVC3,
Metro Style Apps, more. Free future releases when you subscribe now!
http://p.sf.net/sfu/learndevnow-d2d
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to