On cross-validation for MLP model evaluation

Abstract

Cross-validation is a popular technique for model selection and evaluation. The purpose is to provide an estimate of generalization error using mean error over test folds. Typical recommendation is to use ten-fold stratified cross-validation in classification problems. In this paper, we perform a set of experiments to explore the characteristics of cross-validation, when dealing with model evaluation of Multilayer Perceptron neural network. We test two variants of stratification, where the nonstandard one takes into account classwise data density in addition to pure class frequency. Based on computational experiments, many common beliefs are challenged and some interesting conclusions drawn. © 2014 Springer-Verlag Berlin Heidelberg.