[pymvpa] default SVM C parameter

[J.A. Etzel]

Several years ago I played with using different values of c, by a 
rigorous grid-search and arbitrarily. I've never seen optimizing make 
enough of a difference with fMRI data to justify the increased hassle 
and time.

Anyone seen anything different?

Jo



On 9/1/2011 9:03 AM, Mark Lescroart wrote:
> Hello all,
>
> I have two questions about the default C parameter for linear SVM
> classifiers in pymvpa (/ LibSVM).
>
> First, I'd like a little more information about how the parameter is
> chosen by default. The help says that the default value is -1, and
> that "In linear kernel, negative values provide automatic scaling of
> their value according to the norm of the data," but I didn't find
> that particularly helpful (What does "according to the norm of the
> data" mean?)
>
> Second, does anyone have any experience choosing a C parameter by
> more rigorous methods (cross-validated parameter selection)? Does
> choosing an optimal C value make a big difference?
>
> THANKS,
>
> Mark
>
> PS - I'm still using the 0.4 version of pymvpa - has setting of the
> default C parameter changed since then? I'm using the LibSVM
> implementation of the LinearCSVMC classifier on MacOS X 10.5.8.
>
> _______________________________________________ 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