Detection of interdependences in attribute selection

Abstract

A new measure for attribute selection, called GD, is proposed. The GD measure is based on Information Theory and allows to detect the interdependence between attributes. This measure is based on a quadratic form of the Ms distance and a matrix called Trans information Matrix. In order to test the quality of the proposed measure, it is compared with other two feature selection methods, namely Ms distance and Relief algorithms. The comparison is done over 19 datasets along with three different induction algorithms.