Are there any quantitative and justifiable methods to help choose minimum sample size for having a 'good' model from SVM? Logistic regression affords power analysis which provides minimal $n$ to guarantee a low type II error. But I can't tell if SVM can be analyzed similarly.

There are all sorts of extremely vague (and poor) statements like '10 instances per feature' or 'a lot more'. I'm trying to find out if there's more than just experience that can help guides users.

I'd expect there to be a very different calculations than for power analysis but maybe there is a kind of general analysis not dependent on having so many assumptions of the model.

A calculation would be preferable, but a more specific rule of thumb would be helpful too (and explanation or source that would justify that rule of thumb).



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.