scikit learn - sklearn: AUC score for LinearSVC and OneSVM -

one option of svm classifier (svc) probability false default. documentation not does. looking @ libsvm source code, seems sort of cross-validation.

this option not exist linearsvc nor onesvm.

i need calculate auc scores several svm models, including these last two. should calculate auc score using decision_function(x) thresholds?

answering own question.

firstly, common "myth" need probabilities draw roc curve. no, need kind of threshold in model can change. roc curve drawn changing threshold. point of roc curve being, of course, see how model reproducing hypothesis seeing how ordering observations.

in case of svm, there 2 ways see people drawing roc curves them:

1. using distance decision bondary, mentioned in own question
2. using bias term threshold in svm: http://researchgate.net/post/how_can_i_plot_determine_roc_auc_for_svm. in fact, if use svc(probabilities=true) probabilities calculated in manner, using cv, can use draw roc curve. mentioned in link provide, faster if draw roc curve directly varying bias.

i think #2 same #1 if using linear kernel, in own case, because varying bias varying distance in particular case.