[pymvpa] Biased estimates by leave-one-out cross-validations in PyMVPA 2

[Yaroslav Halchenko]

yeah -- we need to improve our documentation of Balancer... meanwhile
try something like

cv=CrossValidation(clf,
                   ChainNode([NFoldPartitioner(),
                              Balancer(attr='targets',
                                       count=1, # for real data > 1
                                       limit='partitions',
                                       apply_selection=True
                                       )],
                             space='partitions'))

On Sat, 21 Apr 2012, Ping-Hui Chiu wrote:

>    Thanks Yaroslav! I tried the Balancer generator but it didn't help in the
>    following case of binary classification on random samples:

>    from mvpa2.suite import *
>    clf=LinearCSVMC();
>    cv=CrossValidation(clf,ChainNode([NFoldPartitioner(),Balancer()],space='partitio
>    ns'))
>    acc=[]
>    for i in range(200):
>     print i
>     ds=Dataset(np.random.rand(200))
>     [1]ds.sa['targets']=np.remainder(range(200),2)
>     [2]ds.sa['chunks']=range(200)
>     results=cv(ds)
>     acc.append(1-np.mean(results))

>    >>>print np.mean(acc),np.std(acc)
>    0.4106 0.212417960634

-- 
=------------------------------------------------------------------=
Keep in touch                                     www.onerussian.com
Yaroslav Halchenko                 www.ohloh.net/accounts/yarikoptic