Fuzzy rough feature selection via stripped decision β-neighborhood set and misclassification ratio

Abstract

Fuzzy β-covering is a type of granular structure for knowledge representation that has been widely used for machine learning and data mining in recent years. In evaluations of the reduction and redundancy of fuzzy β-coverings, most of the existing methods generate a new fuzzy β-covering for describing the similarity between objects to select important fuzzy β-coverings. However, not all fuzzy β-neighborhoods in the generated fuzzy β-covering are necessary for further determining important fuzzy β-coverings. Therefore, in this study, we propose the concept of a stripped decision β-neighborhood set, and present a fuzzy β-covering reduction approach based on the misclassification ratio for feature subset selection. Inspired by the lower approximation operator of rough sets, the concept of a stripped decision β-neighborhood set is first proposed to remove some unnecessary β-neighborhoods for further determining important fuzzy β-coverings. Moreover, the connection between the stripped decision β-neighborhood set and positive region is discussed in a fuzzy β-covering group decision system. The misclassification ratio for fuzzy β-coverings is then defined on this basis. An accurate feature selection method is presented based on the knowledge represented by fuzzy β-coverings and fuzzy β-covering reduction by using the misclassification ratio. Finally, the experimental results demonstrated the effectiveness of our method compared with several other excellent feature selection methods under four classical classifiers.