Neighborhood rough set based ensemble feature selection with cross-class sample granulation

• Published in: Applied soft computing Vol. 131; p. 109747
• Main Authors: Liu, Keyu, Li, Tianrui, Yang, Xibei, Yang, Xin, Liu, Dun
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.12.2022
• Subjects: Attribute reduction; Ensemble feature selection; Granular computing; Neighborhood rough set; Neighborhood rough set; Granular computing; Ensemble feature selection; Attribute reduction
• ISSN: 1568-4946; 1872-9681
• DOI: 10.1016/j.asoc.2022.109747

Exploring feature significance associated with label is a fundamental task in the architecture of feature selection. Nevertheless, most of the existing schemes are limited by the global feature significance over the entire universe. It follows that some specific characteristics of features implied in sample subspaces may be overlapped. To fill such a gap, a novel ensemble feature selection with cross-class sample granulation is developed. Our method explicitly involves two main phases: (1) cross-class sample granulation — data is separated into multiple granules which are generated by querying the locations of samples in their respective classes, so as to provide local bases; (2) ensemble feature selection — localized evaluations of feature significance are integrated which are induced by leveraging multiple homogeneous fine-granularity measures from those bases, so as to select qualified features. To validate the effectiveness of our proposed method, it is compared with several well-established feature selection schemes in CART, KNN and SVM classification performance. Experimental results on 20 UCI data sets demonstrate that our method is superior as it yields higher accuracy with satisfactory elapsed time. •A cross-class sample granulation mechanism is proposed.•A fine-granularity neighborhood approximate quality is introduced for localized feature evaluation.•A novel ensemble feature selection scheme is presented by leveraging multiple homogeneous fine-granularity measures.