Random subspacing for regression ensembles

Abstract

In this work we present a novel approach to ensemble learning for regression models, by combining the ensemble generation technique of random subspace method with the ensemble integration methods of Stacked Regression and Dynamic Selection. We show that for simple regression methods such as global linear regression and nearest neighbours, this is a more effective method than the popular ensemble methods of Bagging and Boosting. We demonstrate that the approach can be effective even when the ensemble size is small.