[pymvpa] Spatiotemporal Feature Selection

Michael Hanke mih at debian.org
Sun Jul 22 08:38:30 UTC 2012


On Thu, Jul 19, 2012 at 04:02:59PM +0200, Roberto Guidotti wrote:
> I've another question about spatiotemporal analysis. In my dataset I have 6
> runs per task with approximately a hundred volumes each task each run and
> running a classification on only 6 examples is very overfitting-oriented,
> how can I do to reduce the overfitting? I've tried with a leave 3 run out
> crossvalidation in order to have more testing runs and a very general model
> fitted only on 3 runs (I use SVM)? Are there some other strategies?

There are many possible strategy (e.g. repeated random selection, ...).
But, again, more information on your dataset would be helpful. You are
saying that you have 100 volumes per that (per run) that you reduce into
a single sample? Why is that?


Michael Hanke

More information about the Pkg-ExpPsy-PyMVPA mailing list