2013/3/25 Mathieu Blondel <math...@mblondel.org>: >> I'm confused. Since when is that so? The other losses definitely support >> OvR multi-class. I would be quite surprised if 'log' does not. > > predict_proba currently raises an exception in the multiclass case: > https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/linear_model/stochastic_gradient.py#L676 > > We can modifiy predict_proba so that it computes the probabilities of > each one-vs-rest task and then normalize them, like we already do > here: > https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/multiclass.py#L111
We've had this discussion several times before; pprett opposes it because it's not well-calibrated. The idea is to either implement multinomial LR (hard to do within the single-output setup of our SGD algorithm) or re-calibrate using isotonic regression. > I'm not sure if there exists any strong theoretical motivation for > doing this normalization but at least this is how liblinear is doing > it and this is also the method recommended by this paper: > http://www.research.ibm.com/people/z/zadrozny/kdd2002-Transf.pdf You mean the score rescaling at the end of section 2? What Zadrozny and Elkan really seem to be recommended is isotonic regression... -- Lars Buitinck Scientific programmer, ILPS University of Amsterdam ------------------------------------------------------------------------------ Everyone hates slow websites. So do we. Make your web apps faster with AppDynamics Download AppDynamics Lite for free today: http://p.sf.net/sfu/appdyn_d2d_mar _______________________________________________ Scikit-learn-general mailing list Scikit-learn-general@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
