Hi, Shreedhar,

Generally, you can try Lasso feature selection as well, it not only do feature selection, but also priorities the features, give you an idea which features are more important than others.

I am not sure SVM is a suitable methods for frequential data process,  but from my experience, sometimes RBF kernel may lead to over-fitting problem, you have to be very careful about the C and gamma selection. 

Hi Shreedhar,

an overview and helpful references on feature selection is given in the article by Liu and Yu: "Toward Integrating Feature Selection Algorithms for Classification and Clustering".

If you want optimal features for your SVM you could use a wrapper method (the term is used in the article from before).  However, when using wrappers overfitting again is a serious issue (see Reunanens work entitled "Overfitting in feature selection: Pitfalls and solutions.")

Greetings, Sebastian

Hi Shreedhar,

It's a very huge topic and has several directions. Anyway, you could apply SVM for attribute selection, and then for classification with the use of suitable evaluation method, e.g., cross-validation.

HTH.

Samer