Dear Harry, All,

WEKA computes a propability distribution for each instance when
performing classification. I couldn't find any documentation in the
manual, here is what I know:

The distribution depends on the classifier, and is supported by some
non-probabilistic classifiers as well (e.g. Nearest Neighbor, the
distribution corresponds to the numbers of neighbors for each class).
Other classifiers that do not support any confidence/probability (e.g.
SMO) simply return 1 for the predicted class value, 0 for the other
class values.

Using the command line you can use -p to output the distribution
instead of the predicted class.

When using the Java API:

Evaluation evaluation = new Evaluation( trainDataset );
evaluation.evaluateModel( classifier, testDataset );
for (int i = 0; i < evaluation.predictions(); i++) {
  NominalPrediction prediction = (NominalPrediction)
evaluation.predictions().elementAt(i);
  double[] distribution = prediction.distribution();
}

Best regards,
Martin


On Tue, Mar 23, 2010 at 2:57 PM, Haralambos Sarimveis
<hsarimv at central.ntua.gr> wrote:
> Hello Martin,
>
>
>
> You mentioned during yesterday’s discussion that all WEKA algorithms can
> produce confidence intervals for the predictions. Can you please give some
> reference?
>
>
>
> Best regards,
>
> Harry
>
> __________ Information from ESET Smart Security, version of virus signature
> database 4967 (20100323) __________
>
> The message was checked by ESET Smart Security.
>
> http://www.eset.com
>



-- 
Dipl-Inf. Martin Gütlein
Phone:
+49 (0)761 203 8442 (office)
+49 (0)177 623 9499 (mobile)
Email:
guetlein at informatik.uni-freiburg.de