Dynamic clustering based on universal gravitation model

Abstract

Hard/fuzzy c-means and agglomerative hierarchical method are famous and useful clustering algorithms. The former are algorithms using "global information of data", the latter is one using "local information of data". One of main results in this paper is proposal for new clustering algorithm (Dynamic Clustering; DC), which has the advantage of them. In DC, clusters are updated by some model introduced in advance. That is, the clusters are moved according to the model, and merged. Here, merging two clusters means that two clusters contact each other. The model is called option of DC. In this paper, two options of DC are proposed, i.e., interaction power model(IP) and unit weight model(UW). Moreover, the way to determine sampling time, which is one of the parameters in DC, is discussed for precise merging of data and shortening of calculation time. Furthermore, through numerical examples, it is shown that DC gives good classification for the data which is difficult to be classified by the former clustering algorithms. © Springer-Verlag Berlin Heidelberg 2005.