A Review on Enhancement to Standard K-Means Clustering

Abstract

In clustering, objects that have similar nature will lie collectively in the same group called cluster, i.e. same cluster and if they are of distinct nature, then they will be belonging to other cluster of similar nature. The standard k-means is a prime and basic procedure of the clustering but it suffers from some shortcomings, these are as follows (1) Its performance depends on initial clusters which are selected randomly in standard k-means. (2) The basic algorithm, standard k-means, has computational time of 0(NKL) where N represents the number of data points, K represents the number of distinct clusters and L is the number of different iterations that is time consuming or too much expensive. (3) The standard k-means algorithm also has the dead unit problems that result in clusters which contains no data points that is empty cluster. (4) In standard k-means if we do random initialization which causes to converse at local minima. Several enhancement techniques were introduced to enhance the efficiency of the basic k-means algorithm but most of the algorithms were focus only on one of the above drawbacks at a time. In this review paper, we consider initial centre as well as computational complexity problem along with dead unit problem in single algorithm.