i am trying to classify a data set with help of the JMySVMLearner. Now i've the following problem:
With GridParameterOptimization i can only find a parameter set for which the classification result for one class ist correct (100%) and for the other class very bad (<=30%). Is it possible to find a parameter set for which you can obtain a balanced classification result (>=75% for each class)?
You can attempt to tilt the SVM learning by wrapping it in a MetaCost operator. In this case you would increase the costs of misclassifying the second class, in the hope that a more balanced performance emerges. Works fine on binominal labels, not confident about polynominals. Also I've found that performance can change quite a bit depending on the correct settings for C and gamma in the libSVM learner.