Component-wisely sparse boosting

Abstract

This paper proposes a gradient boosting method, which provides a component-wisely sparse solution. Here, 'component-wisely sparse' implies that none of base learners associated with certain input variables are not included in the final solution. Our proposed method consists of two major promising results compared to existing standard boosting methods: first, the proposed method makes the interpretation of the estimated model a bit easier since less input variables are used in the estimated model. Second, the proposed model yields better prediction accuracy even when there are many noisy input variables. Also, the computation of the proposed method is almost identical to that of standard boosting methods. Subsequently, it can be easily applied to large data sets. The proposed methodology is illustrated on a simulation study and real data. (C) 2011 The Korean Statistical Society. Published by Elsevier B.V. All rights reserved.