A Study on Initial Centroids Selection for Partitional Clustering Algorithms

Abstract

Data mining tools and techniques allow an organization to make creative decisions and subsequently do proper planning. Clustering is used to determine the objects that are similar in characteristics and group them together. K-means clustering method chooses random cluster centres (initial centroid), one for each centroid, and this is the major weakness of K-means. The performance and quality of K-means strongly depends on the initial guess of centres (centroid). By augmenting K-means with a technique of selecting centroids, several modifications have been suggested in research on clustering. The first two main authors of this paper have also developed three algorithms that unlike K-means do not perform random generation of the initial centres and actually produce same set of initial centroids for the same input data. These developed algorithms are sum of distance clustering (SODC), distance-based clustering algorithm (DBCA) and farthest distributed centroid clustering (FDCC). We present a brief survey of the algorithms available in the research on modification of initial centroids for K-means clustering algorithm and further describe the developed algorithm farthest distributed centroid clustering in this paper. The experimental results carried out show that farthest distributed centroid clustering algorithm produces better quality clusters than the partitional clustering algorithm, agglomerative hierarchical clustering algorithm and the hierarchical partitioning clustering algorithm.