Hi, On Sat, Nov 20, 2010 at 02:21:17PM +0100, Thorsten Kranz wrote: > thank you for your quick and precise reply. I'll try your proposals on > monday. But one question remains for me: what is the adequate way to > group things together in chunks? All samples with the same label in > one chunk? I must say that I didn't really make use of chunks up to > now...
'chunks' are a way to group samples together that cannot be considered independent of each other. For example, in fMRI data brain volume acquired within a small time frame are heavily correlated for "technical" reasons. 'chunks' can be used to group these together to prevent samples from the same chunk to appear in training and testing dataset at the same time, and therefore introducing an opportunistic bias in the classification error. How samples can be grouped into chunks in your dataset depends on the properties of the data. If, for example, all samples are already independent of any other you could create arbitrary groups of samples to change the granularity of a cross-validation analysis -- e.g. always having 10 samples in the test dataset instead of just one. HTH, Michael -- GPG key: 4096R/7FFB9E9B Michael Hanke http://mih.voxindeserto.de _______________________________________________ Pkg-ExpPsy-PyMVPA mailing list [email protected] http://lists.alioth.debian.org/mailman/listinfo/pkg-exppsy-pymvpa
