Pseudolabel Decision-Theoretic Rough Set

Abstract

In decision-theoretic rough set (DTRS), the decision costs are used to generate the thresholds for characterizing the probabilistic approximations. Similar to other rough sets, many generalized DTRS can also be formed by using different binary relations. Nevertheless, it should be noticed that most of the processes for calculating binary relations do not take the labels of samples into account, which may lead to the lower discrimination; for example, samples with different labels are regarded as indistinguishable. To fill such gap, the main contribution of this paper is to propose a pseudolabel strategy for constructing new DTRS. Firstly, a pseudolabel neighborhood relation is presented, which can differentiate samples by not only the neighborhood technique but also the pseudolabels of samples. Immediately, the pseudolabel neighborhood decision-theoretic rough set (PLNDTRS) can be constructed. Secondly, the problem of attribute reduction is explored, which aims to further reduce the PLNDTRS related decision costs. A heuristic algorithm is also designed to find such reduct. Finally, the clustering technique is employed to generate the pseudolabels of samples; the experimental results over 15 UCI data sets tell us that PLNDTRS is superior to DTRS without using pseudolabels because the former can generate lower decision costs. Moreover, the proposed heuristic algorithm is also effective in providing satisfied reducts. This study suggests new trends concerning cost sensitivity problem in rough data analysis.