A methodical study on behavior of different seeds using an iterative technique with evaluation of cluster validity

Abstract

Data analysis methods are vital for analyzing the rising colossal scale of high-dimensional data. Today, cluster analysis is a widely known technique applied and universally practised in many research areas. Among ‘n’ different clustering techniques we briefly deal with centroid model, i.e., K-means, which is an iterative clustering technique. The recital of this algorithm is dependent on certain factors, which include the selection of initial centroid and the approach used in performing reckoning from each data point to different cluster centers. Initial pattern considered randomly by K-means algorithm often make the clustering results reach the local optima, i.e., choice of initial seed (pattern) greatly affects the ultimate clusters that results, in terms of inter and intra cluster distances and firmness. In this research paper author experimented the behaviors of different patterns with different distance metrics on k-means. Finally estimated validity check, i.e., cluster division ratios for every distance measure used and patterns considered. The experimental grades showed the maximum cluster parting and observed better cluster quality when chosen the initial seed as per the assumptions made, compared to patterns randomly picked. The pragmatic results were met when tried with different sample set and different k values. Further an addition of automatic detection of ideal initial pattern as mentioned by author statements leads to an additional trait to k-means.