Supervised information granulation strategy for attribute reduction

In rough set based Granular Computing, neighborhood relation has been widely accepted as one of the most popular approaches for realizing information granulation. Such approach is to group samples in terms of their similarities without the consideration of their labels. Therefore, it can be referred to as an unsupervised information granulation strategy. Nevertheless, it is obvious that such unsupervised mechanism may generate imprecise neighborhoods by comparing the actual labels of samples. It follows that it is not good enough for classification-oriented attribute reduction to select qualified attributes. To fill such a gap, a novel supervised information granulation strategy is proposed. Different from the unsupervised information granulation, samples are grouped by using not only the similarities over conditional attributes but also the labels. For such a purpose, our mechanism mainly contains two aspects: (1) intra-class radius, which aims to add samples with the same label into neighborhood; (2) extra-class radius, which aims to delete samples with different labels from the neighborhood. The experimental results over 12 UCI data sets demonstrate that, compared with previous researches, the reducts derived by our supervised information granulation may contribute to superior classification performances. This study suggests new trends and applications of considering information granulation from the viewpoint of supervised learning.