Estimation of Clusters Number and Initial Centers of K-Means Algorithm Using Watershed Method

In K-means clustering algorithm, the selection of cluster number k and initial K-means center has certain influence on the result. It would generate very different aggregation result when confronting with some certain types of data set. This paper aims at proposing an estimation method to evaluate the initial parameters for K-means algorithm. The estimation is executed through data analysis, which contains two main steps: the data would be transformed into data dimensional density first, and then, watershed method would be applied to divide the data space into multiple regions. Each regional center is selected as an initial K-means center, and the number of region is set as cluster number. This estimation method takes advantage of image segmentation ideology and the case study in this paper showed its favorable performance.