Local Random Subspace Method for Constructing Multiple Decision Stumps

We propose a technique of localized multiple decision stumps. The ensemble consists of multiple decision stumps constructed locally by pseudorandomly selecting subsets of components of the feature vector, that is, decision stumps constructed in randomly chosen subspaces. The idea of the local ensemble is that although no single function works well globally, in any local region a function should be capable of doing the classification. We performed a comparison with other well known combining methods using decision stump as based learner, on standard benchmark datasets and the proposed method gave better accuracy.