Measuring the generalization capabilities of arbitrary classifiers
Mark E. Oxley, Amy L. Magnus · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2003
Given a classifier trained on two-class data one wishes to determine how well the classifier will perform on new, unseen data. To do this task one typically uses the data to estimate a distribution of the data, generate new data from this distribution, and then test the data. Also, hold-out methods are used including cross-validation. We propose a new method that uses computational geometry techniques that produces a partial ordering on subsets in feature space and measures how well the classifier will perform on these subsets. There are some conditions on the classifier that must be satisfied in order that this measure, in fact, exists. We give the details for these conditions as well as the results concerning this special collection of classifiers. We derive the measure that quantifies the generalization capability for the special collection classifier.