Some Versions of K-means Clustering Method and its Comparative Study in Low and High Dimensional Data

Abstract:

In this paper we present some versions of k-means clustering method and compare the methods using simulated data, and also low and high dimensional data set in terms of their accuracy and minimized total intra-cluster variance. The versions of k-means clustering method discussed in this paper are namely: The Forgy’s method, Lloyd’s method, MacQueen’s method, Hartigan and Wong’s method, Likas’ method and Faber’s method. These methods minimize a given criterion by iteratively relocating points between clusters until a locally optimal partition is attained. In a basic iterative algorithm, such as k-means, convergence is local and the globally optimum solution cannot be guaranteed. From experimental results, it was observed that Likas’ method and Faber’s method performed better in our synthetic data; method like Likas’ performed better in low dimensional data (iris data) while Hartigan and Wong’s method did better in high dimensional data (yeast cell cycle data).