Feature Selection for Partially Labeled Data Based on Neighborhood Granulation Measures

Abstract

As an effective feature selection technique, rough set theory plays an important part in machine learning. However, it is only applicable to labeled data. In reality, there are massive partially labeled data in machine learning tasks, such as webpage classification, speech recognition, and text categorization. To effectively remove redundant features of partially labeled data, the neighborhood granulation measures based on a neighborhood rough set model are put forward in this paper, which can be used to evaluate the discernibility ability of feature subsets under both information systems and decision systems. Moreover, a new definition of significance is introduced. Based on that, a semisupervised reduction algorithm is presented for the feature selection of partially labeled data. Several datasets are chosen to verify its effectiveness. The comparative experiments show that our proposed method is more effective and applicable to the feature selection of partially labeled data.