Just a simple question on parameter selection for SVMs. If I use a minimum finding algorithm to find the optimal parameters for a set of data, how do I "average" the parameters over a set of cross validation runs to come up with the best parameters for my test set run? When doing X-fold cross validation, I can see simply averaging the fit of the model for each parameters set, but not sure how to do it with a minimum finding algo.



It is best to think of cross-validation as giving you a performance estimate for a model fitting procedure, which includes selection of the hyper-parameters, in which case the best thing to do would be to use the same procedure to fit an SVM and tune the hyperparameters to the whole training set and use that to make the test predictions.

  • 1
    $\begingroup$ Hey sorry didn't quite understand your answer. In my cross validation, i split my train set into train/test sets, and then run either a grid search of min finding algo on a train set, and optimize parameters according to the test set. in grid search, i can average the fit over the cross validated parameters, but can't in min finding algo. when trying to predict a new test set, i can't optimize on the new test set and optimizing on the whole train set would just overfit the training data. i probably misunderstood your answer though. $\endgroup$ – tomas Feb 8 '12 at 19:34

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.