|
I have been using a logistic regression solver for some prediction task. I am wondering whether or not it is okay to take the same dataset and run SVM on it? what are the assumptions for the data for SVM? should I change my data at all? |
|
Strictly speaking you should have no problem with the data. Be sure the kernel you use is compatible though, with your data, you might need some pre processing, but most "out of the shelf" programs do that automatically. Since you have a problem solved with linear regression, I do not think you will have problems with the convexity of your problem, so you are pretty much good to go if you use libSVM @palafox, could you elaborate more on your statement, "Since you have a problem solved with linear regression, I do not think you will have problems with the convexity of your problem, " why the successful of linear regression can indicate the convexity of problem
(Feb 16 '14 at 15:36)
ouyang
|
What processing are you doing on the data? I would at least do a component-wise normalization of the data in both cases.
The answer is essentially yes. SVM is a discriminative approach. Both it and logistic regression were made for binary classification. They should be interchangeable.
I wish to make a comment. I have tried logistic regression and SVM (libSVM) on two projects related to multi-class text classification. In both cases, logistic regression yields better performance (not too much). For SVM, I tried linear kernel and RBF kernel.