A novel approach for feature selection using Rough Sets

Abstract

Rough Set is a mathematical tool to find patterns hidden in data with uncertainty. A major step for reduction of high dimension data, present in various forms, is selection of appropriate features. In this work we propose a new indiscernibility relation based on clusters, and compare its effectiveness with that of classical Rough Set based indiscernibility. In particular, we study the proposed Rough Set based scheme for feature set reduction. Rough-Cluster (RC) based approximate algorithms are proposed. The major advantage of these algorithms over the classical method is that they work well even without data discretization. The accuracy, measured in terms of the proportion of correctly classified data samples, is obtained on various standard data sets. The results are found to be on par with those obtained through classical Rough Set based technique for the problem of feature selection.