Feature Subset Selection Based on Variable Precision Neighborhood Rough Sets

Abstract

Rough sets have been widely used in the fields of machine learning and feature selection. However, the classical rough sets have the problems of difficultly dealing with real-value data and weakly fault tolerance. In this paper, by introducing a neighborhood rough set model, the values of decision systems are granulated into some condition and decision neighborhood granules. A concept of neighborhood granular swarm is defined in a decision system. Then the sizes of a neighborhood granule and a neighborhood granular swarm are also given. In order to enhance the fault-tolerant ability of classification systems, we define some concepts of granule inclusion, variable precision neighborhood approximation sets and positive region. We propose a variable precision neighborhood rough set model, and analyze its property. Furthermore, based on the positive region of a variable precision neighborhood, we give the significance of an attribute and use it to select feature subsets. A feature subset selection algorithm to the variable precision neighborhood rough sets is designed. Finally, the feature selection algorithm is carried out on the UCI datasets, and the selected features are tested by the support vector machine (SVM) classification algorithm. Theoretical analysis and experiments show that the proposed method can find the effective and compact feature subsets, which have abilities of fault tolerance.

Keywords

• Rough sets
• Variable precision neighborhood rough sets
• Attribute reduction
• Feature selection
• Neighborhood systems