Abstract:Many cluster validity indices faced two problems. First, most of them neglect the variance of intra-cluster compactness between clusters. Second, some indices can not deal well with the clusters of arbitrary shapes. In this paper, we propose a new approach to solving these two problems: we define the validity indices for the whole clustering results upon the validity indices of single clusters. This idea can help indices deal with clusters of different intra-cluster compactness. After that, we take advantage of connect distance and propose a new connectivity-based cluster validity index which can deal with the clusters of arbitrary shapes. The experiment results show its effectiveness.

A New Connectivity-Based Cluster Validity Index