[pymvpa] pymvpa one-class svm

[basile pinsard]

Well as far as I understand, it seems that these are two different problems
(see here
<http://scikit-learn.org/stable/auto_examples/svm/plot_oneclass.html#example-svm-plot-oneclass-py>),
I might be wrong.
One is to find a separating hyperplane between two population of samples
The other is to find border around one population and a measures of
confidence about samples to belong to this distribution.
Rest can be quite versatile and it might be misclassified if it is not a
brain state that is in the training set.

As far as I have explored the 2 options that I listed:
SVM(svm_impl='ONE_CLASS') hangs when doing a train(ds) but if I select only
samples from the one class it does not run forever.
The problem then is that cross-validation scheme will never give only the
one-class subset to train, which makes it hang.
SVMlib might not be able to deal with targets in that case.

SKLLearnerAdapter(OneClassSVM()) does seem to train only with the one-class
samples and uses the others for prediction in cross-validation.

basile


On Wed, Jan 21, 2015 at 5:16 PM, J.A. Etzel <jetzel at artsci.wustl.edu> wrote:

> I'd be tempted to consider task vs. rest a two-class problem - task and
> rest!
>
> Jo
>
>
>
>
> On 1/21/2015 4:12 PM, basile pinsard wrote:
>
>> Hi pyMVPA!
>>
>> does anyone has ever tried to use one-class SVM (or other 1-class
>> classifier) in PyMVPA.
>> I tried using :
>> SVM(svm_impl='ONE_CLASS')
>> but I am not sure what are the targets supposed to be: boolean or {-1,1}
>> I do not know if the wrapper passes to training only the positive
>> samples or not? while in a cross-validation scheme we would test on full
>> dataset to find false-positives.
>> I also tried using sklearn.svm.OneClassSVM:
>> SKLLearnerAdapter(OneClassSVM(kernel='linear'))
>>
>> I am specifically interested in classifying a task vs. rest which is why
>> I aim at one-class.
>>
>> Thanks for your advices or ideas.
>>
>> Regards
>>
>> basile
>>
>>
>> _______________________________________________
>> Pkg-ExpPsy-PyMVPA mailing list
>> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
>> http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa
>>
>>
> --
> Joset A. Etzel, Ph.D.
> Research Analyst
> Cognitive Control & Psychopathology Lab
> Washington University in St. Louis
> http://mvpa.blogspot.com/
>
> _______________________________________________
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa
>



-- 
Basile Pinsard

*PhD candidate, *
Laboratoire d'Imagerie Biomédicale, UMR S 1146 / UMR 7371, Sorbonne
Universités, UPMC, INSERM, CNRS
*Brain-Cognition-Behaviour Doctoral School **, *ED3C*, *UPMC, Sorbonne
Universités
Biomedical Sciences Doctoral School, Faculty of Medicine, Université de
Montréal
CRIUGM, Université de Montréal
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20150122/014ec1f7/attachment.html>