OV crashes when using “Classifier trainer” box with SVM classifier

[FilippoCastellani]

Good Morning and happy new year !

I am using OV version 3.1.0 and I want to use a SVM classifier with a linear kernel to classify Task vs Rest condition with 2 features and ~20 observations per condition, in a custom scenario. However, when I use the "Classifier trainer" box with 10 or 5 times cross validation, when it starts performing the classification, OV sends a warning message "k-fold test may take a long time, be patient" and then crashes, closing.

I also tried to use the "Classifier trainer" box of OV version 2.2.0 which allows to set all classification parameters such as the kernel, OV does the classification but the classification performance and the confusion matrix are weird (cross-validation test accuracy: 37% and training set accuracy: 52%). Also, using a LDA classifier with both the versions of the box manages to achieve 80% accuracy.

Has anyone succeeded in using and getting good classification performances with the SVM on OV? Are there any special requirements for input features when using the SVM classifier?

Best wishes.

[Thibaut]

Hi,
There is a problem with the current version of OV (fixed for the next one) the parameters are not updated and the crash is due to the fact that the SVM has different parameters from the LDA so the crash occurs. It's annoying but if you set your box on OV 2.2.0. The settings will be good and you will be able to use it without problem on OV the latest version of OV.
Thibaut

[FilippoCastellani]

Hi Thibaut, thank you very much for your answer!
We did try to do exactly what you told us:
- created "classifier trainer" box in O.V. 2.2.0
- imported it in O.V. 3.1.0 and used it successfully
- this time it doesn't crashes but the problem concerning the confusion matrix persists

I also tried to use the "Classifier trainer" box of OV version 2.2.0 which allows to set all classification parameters such as the kernel, OV does the classification but the classification performance and the confusion matrix are weird (cross-validation test accuracy: 37% and training set accuracy: 52%). Also, using a LDA classifier with both the versions of the box manages to achieve 80% accuracy.

The problem is not the crash anymore but the very "weird" results we're obtaining.
Could there be a specific reason why LDA performs correctly while SVM does not ? Could this be related to the O.V. version ?

Thank you very much.
Best wishes, Filippo Castellani.