LIBSVM parameter selection

Question

I’m working on a system that can optimize parameters of SVM (LIBSVM toolbox) and best features (inputs) together. My optimization techniques is harmony search or genetic algorithm. I normalized data before insert it to system ( Maxmin or whitening) so as you know I must set

-g gamma : set gamma in kernel function 
-c cost : set the parameter C of C-SVC, epsilon-SVR, and nu-SVR

I have a binary classification problem with financial database. Besides it sometimes I’m using dimension reduction techniques like “PCA” and other techniques.

So what is your proposed searching intervals for gamma and cost parameters?

Second question : What other parameters of LIBSVM do you think i should optimize with my optimization technique?

Thanks.

PS. My kernel function is "RBF"

Answer 1

So what is your proposed searching intervals for gamma and cost parameters?

Basically, you should do a heuristic grid search by taking some educated guesses at the grid cell sizes, hoping to find a good optimum. Take a look in the Grid.py file in the LIBSVM package. It will give you some hints regarding these values. There are more advanced techniques for doing this, but this is the first thing worth trying.

Second question : What other parameters of LIBSVM do you think i should optimize with my optimization technique?

This depends on the kernel that you choose. Note that gamma doesn't impact all kernel types (like, for example, the linear one). There may be other things that you can try to tweak, but start with these two and see how it goes.

Answer 2

I would suggest you looking at the tutorial provided by Dr. Lin Tips and tricks for liblinear . Unless you have a very strong justification of using a specific kernel try with all available kernels. Again, I would try training/testing with un-normalized/unscaled data too. Remember these (scaling) are heuristics, no one knows what would be the set of params for your requirement. Explore all.