Analysis of variant approaches for initial centroid selection in K-means clustering algorithm

Abstract

The basic K-means procedure is modified aimed at an efficient selection of initial seeds. Different final partitions may be formed as a result of different choices of initial seed sets. An iterative partitional grouping procedure is used and the grouping correctness is calculated. In this work variant approaches for finding initial centroids in k-means are proposed. The initial centroids are chosen using these different approaches: (a) Random generation, (b) Buckshot approach, and (c) ranking technique. The research analyzed the influence of the initial seed selection on cluster quality in k-means algorithm with three different similarity measures in synchrony with various vector representations. The initial centroids chosen play a crucial role toward the accuracy of the clusters and efficiency of the partition-based grouping systems. In the experiment, k-means procedure is applied, and also initial centroids for k-means are chosen by using different proposed approaches. Our investigational outcomes display the accuracy in clusters and efficiency of the k-means procedure is improved compared to traditional way for choosing initial centroids. A numeral of trials was performed and the statistical significance of the consequences is ensured using entropy. The applied methods unfolded the clustering performance for Pearson, Cos and Jaccard correlation resemblances.