On Dynamic Clustering and Two Options
Yasunori Endo, H. Iwata · 2005
Hard/fuzzy c-means and agglomerative hierarchical method are representative clustering algorithms. The one is an algorithm using "global information of data", the other is using "local information of data". In this paper, a new clustering algorithm (dynamic clustering; DC) is proposed, which has the advantages over the conventional clustering algorithms. On DC, clusters are updated by some model introduced in advance. That is, the clusters are moved according to the introduced 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, through numerical examples, it is shown that DC gives good classification for the data which is difficult to classify by the former clustering algorithms