Neural Network Ensembles from Training Set Expansions

In this work we propose a new method to create neural network ensembles. Our methodology develops over the conventional technique of bagging, where multiple classifiers are trained using a single training data set by generating multiple bootstrap samples from the training data. We propose a new method of sampling using the k-nearest neighbor density estimates. Our sampling technique gives rise to more variability in the data sets than by bagging. We validate our method by testing on several real data sets and show that our method outperforms