On 04/17/2012 08:56 PM, Olivier Grisel wrote: > Ok so this is a real bug. Please open an issue. I don't really know how to do that !? > Also I am curious: what does it mean to do OneClass SVM with a linear > kernel? I thought OneClass SVM was for density estimation and I don't > see how on could define a finite density function (even if not > normalized) with an hyperplane. In fact I have used to use the libsvm pything interface and where giving a parameter linear or rbf it was not matter so much at least for my data, i.e. Automated Genre Identification for web pages in particular the Santini's corpus. However, since I am trying to do some high scale tests python structures like lists or indices used by this interfce are not sufficient. So I had to go to numpy/scipy/pytables and of course scikits way.
In addition, for more that 10000 feature in practice most of the time linear kernel works better compare to other kernels. I seem that this has to do with the overfishing issues and the data are so sparse due to the huge space of features where they are projected that hyperplane seems sufficient or even better choice. So, I just used this rule of thumb for OneClassSVM. I am not sure thought and everything is under investigation. I will try the lowlevel svm I hope the bug will be fixed soon. regards, Dimitrios ------------------------------------------------------------------------------ Better than sec? Nothing is better than sec when it comes to monitoring Big Data applications. Try Boundary one-second resolution app monitoring today. Free. http://p.sf.net/sfu/Boundary-dev2dev _______________________________________________ Scikit-learn-general mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
