[pymvpa] SVM classification of data with temporal correlation

[Vadim Axel]

Hi,

A simple question, though I am not sure how simple is the answer.
I am using SVM for classifying block design data. This data is clearly
temporally correlated. On average, one TR lag is results 0.3 correlation
(autocorr matlab function). One prof of mine in the university said, that
it's forbidden to use SVM in such a case because of i.i.d SVM assumption.
However, I do not remember  any fMRI classification paper which raised this
issue, though it should be applicable to any raw-data classification. This
is in contrast to GLM analysis in SPM, which does make some non-sperecity
corrections. To minimize the temporal correlation I averaged data points in
my block (6 TRs), so now the temporal correlation between consecutive blocks
is only 0.15, but the pattern of classification remained pretty similar to
the raw data.

What do you think about this issue? Does ignoring temporal correlation may
just decrease the prediction rate or it casts doubt in the results in
general?

Thanks,
Vadim
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20091203/b33c5757/attachment.htm>