It's common in machine learning to train on 70% of non-validation data, testing on the remaining 30%. I'm not sure whether the motive for this is theoretical or "empirically this works well", although my reading suggests the latter. Where was this ratio first justified either way?
Add a comment
|