A fast K-Means clustering using prototypes for initial cluster center selection

Abstract

K-Means clustering is a well studied algorithm in literature because of its linear time and space complexity. K-means clustering algorithm selects the initial seed points randomly. The final cluster results obtained and the speed of convergence of solution depends on the initial seed points selected. In this paper we present leaders community based k-means clustering (lc k-means) algorithm that selects good initial cluster centers for k-means clustering to start with. The proposed algorithm runs in two phases where in the first phase a set of prototypes of original dataset are derived by scanning the entire dataset once. The prototypes are grouped further into communities. Initial seed points are derived from these communities. In the second phase k-means algorithm is run over the prototypes derived in the first phase and once solution is converged the prototypes are replaced by its respective followers. Experimental results show that proposed algorithm is very accurate in detecting well separated clusters and also converges solution faster than traditional k-means algorithm.