Also for some values of sigma you won't have the inequality outliers/n <=
nu <= sv/n or the proportion of outliers will be very far from nu.
These values of sigma are clearly not the ones you want to use.
Le mercredi 10 septembre 2014, Albert Thomas <albertthoma...@gmail.com> a
écrit :
> Hi Roberto,
> I've been working with One-Class SVM lately. You may want to have a look
> at two papers from Tax talking about the issue of tuning the parameter
> $\sigma$ of the gaussian kernel when using one-class classifier. The papers
> are about SVDD but it is equivalent to One-Class SVM. The two papers are
> 1) Uniform object generation for optimizing one class classifiers. Tax and
> Duin
> 2) A consistency-based model selection for one-class classification. Tax
> and Muller
>
> Albert
>
>
>
> 2014-09-10 20:13 GMT+02:00 Pagliari, Roberto <rpagli...@appcomsci.com
> <javascript:_e(%7B%7D,'cvml','rpagli...@appcomsci.com');>>:
>
>> Yes, that's what I thought. I guess I will try to find out somewhat
>> empirically. I don't really know what to use for scoring with grid search,
>> especially because sometimes you only have examples of one class.
>>
>>
>> Thank you,
>>
>>
>> -----Original Message-----
>> From: Andy [mailto:t3k...@gmail.com
>> <javascript:_e(%7B%7D,'cvml','t3k...@gmail.com');>]
>> Sent: Wednesday, September 10, 2014 1:23 PM
>> To: scikit-learn-general@lists.sourceforge.net
>> <javascript:_e(%7B%7D,'cvml','scikit-learn-general@lists.sourceforge.net');>
>> Subject: Re: [Scikit-learn-general] one-class SVM confidence values
>>
>> On 09/10/2014 05:01 PM, Pagliari, Roberto wrote:
>> > Thank you.
>> >
>> > One last question. How do you set the value of gamma? From the original
>> paper, it seems like gamma=0 is not the optimal choice. They suggest to use
>> gamma > 0, but I'm not sure how to choose the v alue.
>> >
>> Err I forgot that OneClassSVM is an unsupervised method.
>> You can try to use GridSearchCV but the problem is that it is unclear
>> what the scoring method should be and how to get ground-truth.
>>
>>
>> ------------------------------------------------------------------------------
>> Want excitement?
>> Manually upgrade your production database.
>> When you want reliability, choose Perforce Perforce version control.
>> Predictably reliable.
>>
>> http://pubads.g.doubleclick.net/gampad/clk?id=157508191&iu=/4140/ostg.clktrk
>> _______________________________________________
>> Scikit-learn-general mailing list
>> Scikit-learn-general@lists.sourceforge.net
>> <javascript:_e(%7B%7D,'cvml','Scikit-learn-general@lists.sourceforge.net');>
>> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
>>
>>
>> ------------------------------------------------------------------------------
>> Want excitement?
>> Manually upgrade your production database.
>> When you want reliability, choose Perforce
>> Perforce version control. Predictably reliable.
>>
>> http://pubads.g.doubleclick.net/gampad/clk?id=157508191&iu=/4140/ostg.clktrk
>> _______________________________________________
>> Scikit-learn-general mailing list
>> Scikit-learn-general@lists.sourceforge.net
>> <javascript:_e(%7B%7D,'cvml','Scikit-learn-general@lists.sourceforge.net');>
>> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
>>
>
>
------------------------------------------------------------------------------
Want excitement?
Manually upgrade your production database.
When you want reliability, choose Perforce
Perforce version control. Predictably reliable.
http://pubads.g.doubleclick.net/gampad/clk?id=157508191&iu=/4140/ostg.clktrk
_______________________________________________
Scikit-learn-general mailing list
Scikit-learn-general@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
