SVM convergence info objective function value - can I use it for feature selection?
2 views (last 30 days)
Hello.., I am trying to select the best features by giving some scores to the features by training the data with support vectors of individual features and looking at the margin or objective function. This is explained in the attached paper. According to that, after training, each single feature combination is associated with a margin value and group of support vectors. Then the feature with "Smallest margin" (or should it be minimum objective function?) is chosen as the best single feature. (page V-799 of the paper).
My question is, for above check, can I use the objective function value that is given in SVM convergence info after running MATLAB fitcsvm ? Usually objective function value is negative. Should I consider the only the magnitude in this case?
Thanks so much Greatly appreciate your time Regards janaka
Ilya on 24 Nov 2014
There is no paper attached, and even if there was, it's unlikely I would read it. What you want to use as an indicator of the feature set quality is up to you. (This is not a MATLAB question.) I just want to confirm that you can use the Objective field from the ConvergenceInfo property to get the value of the objective function. You can also use the resubMargin method to get classification margins for all observations in the training data (of which you can always find the smallest).