LibSVM biased?

WizMan

Hi, I'm trying to train a model with multiple classes (text mining) and I'm using the LibSVMLearner with a linear kernel type. The problem is that it always creates 2 support vectors for the last class (whichever that may be). This makes it very biased towards that class so that anything I apply the model on is classified as that last class.
How can I avoid this? I'm quite new to the program and data mining so I may have just missed something obvious.

Thanks in advance!

IngoRM

Hi,

I am not sure if I got the point but it could be that the setup is not correct since I never noticed anything like this myself. You could try for example the JMySVMLearner inside of a Binary2MultiClassLearner instead of the SVM and check if this also produces your problem. You could also post your process and if possible even sample data here so we could have a look at the process and see if we find any obvious problems.

Cheers,
Ingo