A new initialization method for clustering categorical data

Abstract

Performance of partitional clustering algorithms which converges to numerous local minima highly depends on initial cluster centers. This paper presents an initialization method which can be implemented to partitional clustering algorithms for categorical data sets with minimizing the numerical objective function. Experimental results show that the new initialization method is more efficient and stabler than the traditional one and can be implemented to large data sets for its linear time complexity. © Springer-Verlag Berlin Heidelberg 2007.