Comparative Assessment of Some Selected Methods of Determining the Number of Clusters in a Data Set
Jerzy Korzeniewski · University of Lodz Repository (University of Łódź) · 2007
This paper is an attempt to compare the performance of an algorithm for determining the number of clusters in a data set proposed by the author with other methods of determining the number of clusters. The idea of the new algorithm is based on the comparison of pseudo cumulative distribution functions of a certain random variable. For a fixed window size we draw К different points and for every point we find the corresponding limiting point in the mean shift procedure. Then we check if the distance (e.g. Euclidean) between every pair of the limiting points is greater than the window size. Analogously we determine the pseudo cumulative distribution functions for different numbers К of clusters. Out of all pseudo cumulative distribution functions we pick the proper one i.e. the last one” (with respect to K) which has a horizontal phase. Other methods of determining the number of clusters in a data set are compared with the proposed algorithm in a number of examples of two dimensional data sets for different clustering methods (k-means clustering and minimum distance agglomeration).