Hi all,
I'm doing classification with CSP components and I'm trying the
support vector machine (svc) with linear kernel and svc with rbf
kernel.
I'm getting always better accuracy with svc with linear kernel than
with rbf kernel.
I expected that the svc with rbf kernel should give better accuracy.
What's the reason?
Why would you expect the rbf kernel to be better? In MEG/EEG linear models
are hard to beat as many effects are either linear in the magnetic field
change or linear in logarithmic power changes. For speed you may even
consider logistic regression which is most of the time equivalent to SVM
but has a true probability model.