Fuzzy rough set attribute reduction based on decision ball model

Abstract

Attribute reduction is a crucial step in data preprocessing in the field of data mining. Accurate measurement of the classification ability of attribute sets stands a central issue in attribute reduction research. The existing fuzzy rough set attribute reduction algorithms measure the classification ability of attribute sets by evaluating the proximity between fuzzy similarity classes and decision classes. However, the granularity of the decision class is too large to reflect the data distribution within the decision class, which may lead to misclassification of samples, thus affecting the effectiveness of attribute reduction. To address this problem, we refine the decision class to propose the concept of decision ball, and study a new extended fuzzy rough set model based on decision ball. In this model, decision balls serve as the evaluation granularity, facilitating the fitting of data distributions and measuring the classification ability of attributes. Expanding on this foundation, we have designed a fuzzy rough set attribute reduction algorithm based on decision ball model (DBFRS). We conducted extensive comparative experiments involving 9 state-of-the-art attribute reduction algorithms on 18 public datasets. Experimental results demonstrate that DBFRS attains high classification accuracy. Moreover, DBFRS exhibits better reduction performance on large and high-dimensional datasets. Compared to current fuzzy rough set methods, DBFRS demonstrates better applicability.