Yes, I met the same problem with you. (see Incosistency of results in
ten-fold cross validation in WEKA
I just read the source code in weka.core.Instances, including the methods
*testCV()*, *trainCV()*, *randomize()*, *stratify()*. I still cannot figure
out the reason that the difference between the two ways.
Sent from: http://weka.8497.n7.nabble.com/