Smoothed emphasis for boosting ensembles

Abstract

Real AdaBoost ensembles have exceptional capabilities for successfully solving classification problems. This characteristic comes from progressively constructing learners paying more attention to samples that are difficult to be classified. However, the corresponding emphasis can be excessive. In particular, when the problem to solve is asymmetric or includes imbalanced outliers, even the previously proposed modifications of the basic algorithm are not as effective as desired. In this paper, we introduce a simple modification which uses the neighborhood concept to reduce the above drawbacks. Experimental results confirm the potential of the proposed scheme. The main conclusions of our work and some suggestions for further research along this line close the paper. © 2013 Springer-Verlag Berlin Heidelberg.