P-k-means: k-means Using Partition Based Cluster Initialization Method

Abstract

The k-Means algorithm is extensively used in a number of data clustering applications. In basic k-means, initial cluster centroids are selected on random basis. As a result, every run of k-means leads to the formation of different clusters. Hence, accuracy and performance of k-means is susceptible to the selection of initial cluster centroids. Therefore, careful initialization of cluster centroids plays a major role on accuracy and performance of the k-means algorithm. In view of this, a new k-means using Partition based Cluster Initialization method called as ‘P-k-means’ is proposed in this paper. The experiment is carried out on six different datasets. The empirical results are compared using various external and internal clustering validation measures. The comparative results show that P-k-means is better than basic k-means.