SVM class doesnt correspond with class probability

KNIME Analytics Platform
1

Hi,

I have a probem: sometimes when using LibSVM classifier for two-class problems I get odd restults, eg:

 

Prob Active | Prob Not Active | SVM Class

0.920 | 0.079 | Not active

 

High Prob Active suggest that compound should have class Active, while it has calculated class Not active.

 

Is it ok?

 

best wishes,

Filip

2

That sounds odd, I haven't seen this happen before.  Can you post an example workflow demonstrating this behavior?

 

 

3

Hi,

(un)fortunatelly I can't reproduce the problem. Maybe it is becouse I've upgraded to KNIME 2.8. Nevertheless I will have my eyes open and report any inconsistency.

Filip

4

Hi,

I still have this problem from time to time. Unfortunatelly I can't create a simple workflow to reproduce the results as data is the property of the company.

zip-file

On this example, the probability to be BBB+ class is 1, while the winner class is BBB- (marked red).