Attribute reduction methods in fuzzy rough set theory: An overview, comparative experiments, and new directions

Fuzzy rough set theory is a powerful tool to deal with uncertainty information, which has been successfully applied to the fields of attribute reduction, rule extraction, classification tree induction, etc. In order to comprehensively investigate attribute reduction methods in fuzzy rough set theory, this paper first briefly reviews the related concepts of fuzzy rough set theory. Then, all methods are summarized through six different aspects including data sources, preprocessing methods, fuzzy similarity metrics, fuzzy operations, reduction rules, and evaluation methods. Among them, reduction rules are reviewed in three categories, i.e., fuzzy dependency-based, fuzzy uncertainty measure-based, and fuzzy discernibility matrix-based. These three types of reduction rules are compared and analyzed through experiments. The experimental results clarify that these three reduction rules can retain fewer attributes and improve or maintain the classification accuracy of a classifier. Moreover, the statistical hypothesis test is conducted to evaluate the statistical difference of these methods. The results show that these algorithms are statistically significantly different. Finally, some new research directions are discussed.