Modified differential evolution based 0/1 clustering for classification of data points: Using modified new point symmetry based distance and dynamically controlled parameters

Abstract

Identification of Clusters is a complex task as clusters found in the data sets are of arbitrary shapes and sizes. The task becomes challenging as identification of all the clusters from a single data set requires use of different types of algorithms based on different distance measures. Symmetry is a commonly used property of objects. Many of the clusters present in a data set can be identified using some point symmetry based distances. Point symmetry based and Euclidean distance measures are individually best in identifying clusters in some particular cases but not together. This article proposes a solution after analyzing and removing the shortcomings in both types of distance measures and then merging the improved versions into one to get the best of both of them. Introduction of differential evolution based optimization technique with dynamic parameter selection further enhances the quality of results. In this paper the existing point symmetry based distance is modified and is also enabled to correctly classify clusters based on Euclidean distance without making a dynamic switch between the methods. This helps the proposed clustering technique to give a speed up in computation process. The efficiency of the algorithm is established by analyzing the results obtained on 2 diversified test data sets. With the objective of highlighting the improvements achieved by our proposed algorithm, we compare its results with the results of algorithm based purely on Euclidean Distance, new point symmetry distance and the proposed modified new point symmetry based distance.