Representative-based classification through covering-based neighborhood rough sets

Abstract

Considerable progress has been made in the theory of covering-based rough sets. However, there has been a lack of research on their application to classification tasks, especially for nominal data. In this paper, we propose a representative-based classification approach for nominal data using covering-based rough sets. The classifier training task considers three issues. First, we define the neighborhood of an instance. The size of the neighborhood is determined by a similarity threshold ?. Second, we determine the maximal neighborhood of each instance in the positive region by computing its minimal ? value. These neighborhoods form a covering of the positive region. Third, we employ two covering reduction techniques to select a minimal set of instances called representatives. To classify a new instance, we compute its similarity with each representative. The similarity and minimal ? of the representative determine the distance. Representatives with the minimal distance are employed to obtain the class label. Experimental results on different datasets indicate that the classifier is comparable with or better than the ID3, C4.5, NEC, and NCR algorithms.