No, they shouldn't be viewed that way.
On Aug 31, 2005, at 5:41 PM, yee seng chan wrote:
Eibe, thanks for the prompt reply! I've 1 more
The reason is that the old version used the
standard sigmoid function
(i.e. without any fitting) to transform the SVM's output into [0,1].
sigmoid function is not parameterized (i.e. not fitted to the
training data). In this case, can the real numbers output from this
old SMO version be considered to be proper probability estimates?
> default mode, the new version computes a "probability" based on the
> number of votes each class receives in pairwise classification. To get
> proper probability estimates you need to turn on the logistic model
> option, which fits a logistic model (i.e. parametrized sigmoid) to the
> output of each SVM in the pairwise classifier and then performs
> pairwise coupling.
> On Aug 31, 2005, at 3:32 AM, yee seng chan wrote:
> > Hi,
> > When executed on the same train.arff and eval.arff, I obtained very
> > different results for 2 different versions of weka SMO. Below are
> > commands:
> > java -mx500m -cp /home/grad/chanys/sw/weka-3-4-5/weka.jar
> > weka.classifiers.meta.MultiClassClassifier -t train.arff -T
> > -p 1 -W weka.classifiers.functions.SMO
> > java -mx500m -cp /home/grad/chanys/wsd/sw/weka-3-2
> > weka.classifiers.MultiClassClassifier -t train.arff -T eval.arff -p
> > -W weka.classifiers.SMO
> > On the older version weka-3-2, the real-numbers returned for each
> > instance is averaging 0.7, but for the newer weka-3-4-5, the
> > real-numbers returned for each test instance is 1.0 (i.e. 100%
> > confident of belonging to a certain class!).
> > What's the difference between these 2 versions? The reason why I
> > to switch to the newer version is because it can fit logistic models
> > to the SVM outputs.
> > Yee Seng._______________________________________________
> > Wekalist mailing list
> > Wekalist(a)list.scms.waikato.ac.nz
> > https://list.scms.waikato.ac.nz/mailman/listinfo/wekalist