Supervised Neighborhood Based Ensemble Attribute Reduction

Abstract

In neighborhood based attribute reduction, neighborhood relation is a typical tool for distinguishing samples. Notably, the neighborhood relation may be powerless in providing satisfactory distinguishing ability. In view of this, the supervised neighborhood based attribute reduction has been explored. However, the supervised neighborhood based reduct may be lack of universality. To file such gap, an ensemble strategy for computing supervised neighborhood based reduct is proposed in our paper. Such ensemble strategy is realized through considering the requirement of each decision class. The experimental results on 8 UCI data sets show that the supervised neighborhood based ensemble strategy can generate reduct not only with higher generalization performance but also with higher stability.