A feature set measure based on Relief

Abstract

s: Feature selection methods try to find a subset of the available features to improve the application of a learning algorithm. Many methods are based on searching a feature set that optimizes some evaluation function. On the other side, feature set estimators evaluate features individually. Relief is a well known and good feature set estimator. While being usually faster feature estimators have some disadvantages. Based on Relief ideas, we propose a feature set measure that can be used to evaluate the feature sets in a search process. We show how the proposed measure can help guiding the search process, as well as selecting the most appropriate feature set. The new measure is compared with a consistency measure, and the highly reputed wrapper approach.