Difference between book and developer version

_______________________________________________
Wekalist mailing list
Wekalist <at> list.scms.waikato.ac.nz
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: Difference between book and developer version

> What precisely is the difference between Weka Book and Developer version?

Read FAQ "What's the difference between book and developer version?". Link
to the FAQs is available from the Weka homepage.

Cheers, Peter
--

-- 
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/           Ph. +64 (7) 858-5174

Plans on parallel WEKA

_______________________________________________
Wekalist mailing list
Wekalist <at> list.scms.waikato.ac.nz
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: CFSS

On Mon, Dec 31, 2007 at 12:35:41PM +0530, Krishna wrote:
> 
>    i need to select best features using this method for classification.
>    what is the percentage cutoff(no of folds ) to select features for
>    classification?

This will vary from problem to problem and is up to the practitioner to
decide, either using any domain knowledge they might have or through
experimentation with a learning algorithm that is to use the selected
features.

Cheers,
Mark.

--

-- 
Mark Hall
Senior Developer/Consultant
Pentaho
The Open Source Business Intelligence Company
Citadel International, Suite 340, 5950 Hazeltine National Dr., Orlando, FL 32822, USA

5 Grace Avenue, Hamilton 3210, New Zealand
+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>

_______________________________________________
Wekalist mailing list
Wekalist <at> list.scms.waikato.ac.nz
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: Two questions about cost sensitive learning

On Sun, Dec 30, 2007 at 01:05:42PM +0800, ?? ?? wrote:
> Hello everyone!
>   My database has 10000 samples of type A and about
> 1000 samples of type B, Moreover type B is more
> important than A. It is both class imbalance and cost
> sensitive. 
>   In weka, as I know, there are
> "CostSensitiveClassifier" and "MetaCost" to solve cost
> sensitive problems, but no method to treat class
> imbalance problems. 
>   The first question is how to treat class imbalance
> problems in weka and whether "CostSensitiveClassifier"
> or "MetaCost" method can be used to solve class
> imbalance problems.

You can use CostSensitiveClassifier in cost sensitive learning
mode to address class imbalances. Depending on the base classifier,
either re-weighting or re-sampling will be used. Note that it uses
an ad hoc heuristic for this when there are more than two classes
in the data.

>   The second question is what is the differences
> between "CostSensitiveClassifier" and "MetaCost".

CostSensitiveClassifier implements cost sensitive learning and cost
sensitive classification (via the minimum expected cost
method). MetaCost is the method by Domingos that achieves a
comprehensible cost sensitive classifier from a bagging approach by
relabeling the training data.

Cheers,
Mark.

--

-- 
Mark Hall
Senior Developer/Consultant
Pentaho
The Open Source Business Intelligence Company
Citadel International, Suite 340, 5950 Hazeltine National Dr., Orlando, FL 32822, USA

5 Grace Avenue, Hamilton 3210, New Zealand
+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>

_______________________________________________
Wekalist mailing list
Wekalist <at> list.scms.waikato.ac.nz
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

controlling cross validation and outputting predictions.

_______________________________________________
Wekalist mailing list
Wekalist <at> list.scms.waikato.ac.nz
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

JRIP's minNo

_______________________________________________
Wekalist mailing list
Wekalist <at> list.scms.waikato.ac.nz
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

Re: controlling cross validation and outputting predictions.

> I want to be able to return the class prediction and know for which instance it 
> relates during cross validation.
> I think i therefore need to control the cross validation step.
> 
> Say, for example, I have 1000 data records and I want to do a 5-fold CV. I want 
> to take records from 1 to 200 as fold1, 201-400 as fold2 , ..., 801-1000 as fold5.
> 
> I have read that this may be possible by editting the following method of the 
> weka.classifiers.Evaluation class:
>   crossValidateModel(Classifier,Instances,int,Random)
> 
> by commenting out the randomize and stratify method calls.
> 
> This potentially will only work on the command line version. I need to see the 
> predictions in the output but can only do this with the gui. On the command line 
> -p 0 only shows the predictions on the training dataset when using cross validation.
> 
> Does anyone know of a solution for identifying  predictions and the instance 
> number for the test data in cross validation?

If you wanna track instances, then use an ID attribute in combination 
with the FilteredClassifier, your base classifier of choice and the 
Remove filter (to get rid of the ID attribute, before the base 
classifier sees the data). See FAQ "How do I use ID attributes?" for 
more information. Link to the FAQs is available form the Weka homepage.

On the other hand, if you wanna have more control over what's happening, 
have a look at the article "Generating cross-validation folds (Java 
approach)":
http://weka.sourceforge.net/wiki/index.php/Generating_cross-validation_folds_%28Java_approach%29

The source file "CrossValidationAddPrediction.java" (in the "Links" 
section) runs a normal cross-validation, but also adds the 
prediction/distribution/error flag to the test data of each fold. At the 
end, the evaluation summary and the "enriched" test data is printed to 
the console.

HTH

Cheers, Peter
--

-- 
Peter Reutemann, Dept. of Computer Science, University of Waikato, NZ
http://www.cs.waikato.ac.nz/~fracpete/     +64 (7) 838-4466 Ext. 5174

Re: controlling cross validation and outputting predictions.

_______________________________________________
Wekalist mailing list
Wekalist <at> list.scms.waikato.ac.nz
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist

BUG in InfoGainSplitCrit.java

Dear WEKAer,

Please refer to
	weka.classifiers.trees.j48.GainRatioSplitCrit

The following method :
///////////////////////////////////////////////////////////////////////////////////
  /**
   * This method is a straightforward implementation of the information
   * gain criterion for the given distribution.
   */
  public final double splitCritValue(Distribution bags) {

    double numerator;

    numerator = oldEnt(bags)-newEnt(bags);

    // Splits with no gain are useless.
    if (Utils.eq(numerator,0))
      return Double.MAX_VALUE;

    // We take the reciprocal value because we want to minimize the
    // splitting criterion's value.
    return bags.total()/numerator;
  }
///////////////////////////////////////////////////////////////////////////////////

There is a BUG here.
The code should be changed into:

    // Splits with no gain are useless.
    if (Utils.eq(numerator,0))
      return 0;

    // We take the reciprocal value because we want to minimize the
    // splitting criterion's value.
    return numerator/bags.total();

Am I right?

This BUG is very harmful to users with experiments by J48.

Best Regards,
Yang ZHANG
------------------------------------------------------------
School of Information Technology and Electrical Engineering
The University of Queensland
Email: mokuram <at> itee.uq.edu.au

_______________________________________________
Wekalist mailing list
Wekalist <at> list.scms.waikato.ac.nz
https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist