I like to find the weight vector for input-space features in a structured SVM. The idea is to identify the most important set of input-space features (based on the magnitude of their corresponding weights). I know that in a binary SVM the weight vector can be written as a linear combination of examples, and the magnitude of those weights represents how much they were effective for the prediction problem at hand. But how do you compute the same for an SSVM?
Asked
Active
Viewed 241 times
3
-
Can you clarify? It sounds like you are looking for feature importance, but the link you give talks about writing weights as a linear combination of input, which is not the same thing at all. – Sean Owen Aug 14 '14 at 10:22
-
@SeanOwen, actually, what I wanted to say was since you can compute the weight vector using the dual variables and the input examples (in a binary SVM), and the actual weights corresponding to features tell us the relative importance of the features Gene Selection for Cancer Classification using Support Vector Machines – imk Aug 14 '14 at 14:48