I was trying to use RBF kernel for a 3 way classification of data. However, the problem is that when I use RBF kernel all the examples are classified as one class. Here is the confusion matrix for RBF kernel
[[ 0 0 19] [ 0 0 12] [ 0 0 19]]
But when I use linear kernel i get a decent accuracy of 80%. I tried the suggestions mentioned in this answer but none of them helped. The dataset contains 500 samples and I have used 3 features. T always thought that when the features are lesser than number of samples rbf will give the best performance but this doesn't seem the case here.