Hi Damien.
Sorry for this apparent regression. There where quite a lot of changes
in the SVM / Logistic regression code, so I can't say directly what the
problem may be.
What I can say is that logistic regression never was multinomial
logistic regression.
It simply wraps LibLinears methods, which implements OvA.
My first guess would be that the problem you are facing is a result of
changing
the meaning of the "C" parameter. It is not invariant to the number of
samples
in the training set, which it is not in LibSVM. So please try and multiply
your "C" with the number of samples, to get the same result.
Hopefully you will have seen a warning that the meaning of C has changed.
If not, we definitely have to work on our warnings / api changes.
By the way, you might be interested in using SGDClassifier for this kind
of data.
Hopefully in the (near) future we will provide multinomial logistic
regression for that.
Cheers,
Andy
On 02/20/2012 08:10 PM, Damien Nouvel wrote:
I recently (yesterday) switched from scikit 0.9 to scikit 0.11 (dev)
so as to get latest code. I use scikit for multiclass classification
from which I retrieve probability estimates. Untli now, I was very
satisfied using LogisticRegression classifier.
Apparently, the LogisticRegression doesn't implement multiclass
classificaiton in a native way anymore? Since I switched, I get very
very poor results (as was the case for SVM previously).
My classification problem is quite high dimensional and very sparse
(45K features * 1M samples). I was suspecting OvA was rather
unefficient for that kind of problem. Could you confirm you moved
LogisticRegression from native multiclass classification to OvA
approach ? Or are there any other changes from 0.9 to 0.11 that may
explain this degradation of performances ?
Regards,
Damien
--
[email protected]
GSM: +33 (0)6 63 56 27 17
------------------------------------------------------------------------------
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
------------------------------------------------------------------------------
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
