What you need to use is the WrapperSubsetEval. This implements the wrapper
approach to subset selection (i.e. repeated 5-fold cross validation on the
training data with respect to a Classifier in order to measure the goodness
of a given subset).
The "cross-validation" option in the "Attribute selection mode" part
the tab is a global evaluation option. When this is selected, the entire
attribute selection process is repeated on each of the X folds of the
data. E.g. there would be X separate runs of WrapperSubsetEval+GreedyStepwise,
each resulting in a separate (probably different) set of attributes. The output
is, for each attribute in the data, how many folds resulted in that attribute
being selected as part of the final subset. I.e. it gives you an indication of
the stability of the feature selection process.
On Tue, Feb 26, 2008 at 06:06:23PM +0100, 'Thomas' wrote:
I cannot understand what's the meaning of the
Cross-validation option in
the "select attributes" tab of the Explorer.
What I want to do is: a Backward elimination (with GreedyStepwise)
evaluated with a classifier (ClassifierSubsetEval); i.e. I want to try
eliminating one attribute per time, then perform a cross-validation, and
if the accuracy is better, permanently eliminate that attribute and
proceed with another elimination.
BUT: in the ClassifierSubsetEval Editor I can only choose beetwen holding
out a test set or using training set.
The Cross-validation option is not in the classifier editor but directly
in the "selct attributes" tab. And this is what I don't understand.
What the meaning of a cross-validation evaluated with, for example, with
Doesn't it have sense only with Classifier Evaluation?
So why it's an external option, and not in its editor?
For my purpose (I've axplained it just above), should i check the
Cross-validation option, and choose "use training" in the
: fai nuove amicizie e ... guadagna con loro grazie a
Wekalist mailing list
Senior Developer/Consultant, Pentaho Open Source Business Intelligence
Citadel International, Suite 340, 5950 Hazeltine National Dr., Orlando, FL 32822, USA
+64 7 847-3537 office, +64 21 399-132 mobile, +1 815 550-8637 fax,
Skype: mark.andrew.hall, Yahoo: mark_andrew_hall
Download the latest release today <http://www.sourceforge.net/projects/pentaho>