Improving clustering and classification performance through effective attribute selection

Abstract

This thesis proposes several novel attribute selection methods for enhancing the performance of Classification and Clustering Algorithms A novel variant of Relief Algorithm namely ReliefS is proposed to weight the relevance of features A new instance reduction algorithm based on Euclidean Distance is proposed to reduce the noisy instances The issue of handling continuous attributes has been tackled through the proposed variant of Fast Correlation based Filter namely FCBF which improves the classification accuracy significantly due to the added discretization step The issue of the elimination of redundant features is handled through the proposed variant of Fast Correlation based Filter namely FCBF which takes into account a novel search strategy that leads to a more balanced attribute elimination An effective Correlation measure called the Combined Symmetrical Uncertainty is employed in the proposed FCBF algorithm which leads to an enhanced redundancy filtering performance The Bayes error based Filter Wrapper hybrid feature selection has been proposed to provide promising learning performance since Bayes error could approximate the search for an optimal newlinesubset of features which provide the least classification error theoretically newline