Hi Roberto,
One possible way to tune the hyperparameters of the One Class SVM is to
split the data set in training and test sets, train the One Class SVM with
the training set and a pre-specified nu, and see if you get a
similar amount of proportion of outliers (a number close to nu) on the test
set. I think that this is what they mean in Libsvm.
But sometimes you can get the same amount of outliers with the training and
test sets but the set returned by the One Class SVM is not the set you are
looking for.
Albert
Le vendredi 5 septembre 2014, Pagliari, Roberto <[email protected]> a
écrit :
> I guess my question was mostly about how gridsearch works with oneclass
> SVM since once-class SVM does not take into account for labels
>
>
>
> Thank you,
>
>
>
>
>
> *From:* Pagliari, Roberto [mailto:[email protected]
> <javascript:_e(%7B%7D,'cvml','[email protected]');>]
> *Sent:* Friday, September 05, 2014 1:36 PM
> *To:* [email protected]
> <javascript:_e(%7B%7D,'cvml','[email protected]');>
> *Subject:* [Scikit-learn-general] one-class SVM with limited number of
> samples
>
>
>
> I am trying to train the one-class SVM.
>
>
>
> According to libSVM
>
> *Q: How do I choose parameters for one-class SVM as training data are in
> only one class? *
>
> *You have pre-specified true positive rate in mind and then search for
> parameters which achieve similar cross-validation accuracy.*
>
>
>
> But the one-class SVM does not take the labels as an input, as it only
> works on X_train.
>
>
>
> 1. So how should cross-validation be performed, assuming the number
> of samples with label ‘1’ is extremely limited.
>
> 2. And what if there is none?
>
>
>
>
>
> Thanks,
>
>
>
------------------------------------------------------------------------------
Slashdot TV.
Video for Nerds. Stuff that matters.
http://tv.slashdot.org/
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
