On 03/14/2012 10:50 PM, Alexandre Gramfort wrote:
> to break ties an idea would be to select the parameters that lead to
> lowest variance i.e. not just look at the mean score across folds
>

Unfortunately the estimated variance across the folds is not a
reliable estimate of the true variance because the classifiers
across the folds - so their scores - are correlated to a degree
which is not easy to estimate. See for example:
http://jmlr.csail.mit.edu/papers/v5/grandvalet04a.html

I am quite interested in this topic so if someone has interesting
references/pointer please share!

Best,

Emanuele


------------------------------------------------------------------------------
This SF email is sponsosred by:
Try Windows Azure free for 90 days Click Here 
http://p.sf.net/sfu/sfd2d-msazure
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to