An adaptive method to determine the number of clusters in clustering process
Huan Doan, Dinh Thuan Nguyen · 2014
A difficult problem of most clustering algorithms is how to specify the appropriate number of clusters. This paper proposes an adaptive method of selecting of number of clusters in clustering process by making coefficients indicated the appropriate number of clusters. The intra-cluster coefficient reflects intra distortion of cluster through maximum distance and a mean distance of cluster's extremely marginal objects. The inter-cluster coefficient reflects distance among clusters. It is ratio between closest distance from this cluster's centre to an extremely marginal object of other cluster and mean distance from this cluster's centre to all of extremely marginal object of other cluster respectively. A new coefficient that indicates the appropriate number of clusters is build from the intra-cluster coefficient and inter-cluster coefficient. The looking for extremely marginal objects and the new coefficient are integrated in a weighted FCM algorithm and it is calculated adaptively while the weighted FCM is processing. The weighted FCM algorithm integrated new coefficient is called FCM++. We experiment with FCM++ on some data sets of UCI: Iris, Wine, Soybean-small and show encouraging results.