Novel way of finding initial means in k-means clustering and validation using WEKA

Abstract

The work proposes a novel choice for the randomly chosen initial means in the k-means clustering. The dataset used for the implementations and validations is the Iris flowers dataset, which contains 150 labeled instances on 5 attributes of the three Iris species. In the k-means clustering, to find the proposed initial means, certain objects are found and eliminated in the clustering, which are very far away from the rest of the objects in their respective clusters. The centroid values of these reduced k clusters are then taken as the initial means in the k-means clustering. The results have shown that the number of iterations required by the algorithm is significantly lesser using the proposed initial chosen means.