An effective algorithm for computing reducts in decision tables

Abstract

Attribute reduction is one important part researched in rough set theory. A reduct from a decision table is a minimal subset of the conditional attributes which provide the same information for classification purposes as the entire set of available attributes. The classification task for the high dimensional decision table could be solved faster if a reduct, instead of the original whole set of attributes, is used. In this paper, we propose a reduct computing algorithm using attribute clustering. The proposed algorithm works in three main stages. In the first stage, irrelevant attributes are eliminated. In the second stage relevant attributes are divided into appropriately selected number of clusters by Partitioning Around Medoids (PAM) clustering method integrated with a special metric in attribute space which is the normalized variation of information. In the third stage, the representative attribute from each cluster is selected that is the most class-related. The selected attributes form the approximate reduct. The proposed algorithm is implemented and experimented. The experimental results show that the proposed algorithm is capable of computing approximate reduct with small size and high classification accuracy, when the number of clusters used to group the attributes is appropriately selected.