Problems for choosing SVR or SVC in LIBSVM?

lucky_q

Dear all,

I want to solve a  simple text classification with only two classes. Obviously C-SVC in libsvm is a OK choice. It's just when I went through some relative papers, I found that the SV regression may also be used to solve these kinds of text classification problems and the results may be more accurate. I got confused.

Could the SV regression really be also chosen to solve classification (eg, text classification) problems?
If so, how could choose the proper parameter to do that?

PS: the training set has already been separated in to two folders appropriately.

I know my questions are quite silly, but I'm really confused.  :-[

Thanks in advance.

IngoRM

Hi,

basically all regression schemes can also be used for classification tasks, this is nothing specific to SVM. In fact, the classification of a linear regression is directly employing this idea, the same is true for the classification meta learner "Classification by Regression". One of the ideas people often have is to directly model the prediction probability instead of using and / or transforming the confidences.

About the optimization of parameters in this setting: Just use a regression performance criterion or transform the predicted regression values into a class before calculating the performance.

Cheers,
Ingo

lucky_q

Hi,

Many many thanks, Ingo. That's very helpful. I guess I still have a lot to learn. Thanks, again.

lucky.