Proposal of Must-Link Constrained K-means with Dynamic Generation of Subordinate Clusters

Abstract

This paper proposes to extend must-link constrained K-means clustering by introducing dynamic generation of subordinate clusters. When clustering high-dimensional data there is a case where data which should belong to the same cluster form several distinct groups in a data space. In order to handle such a case without using distance metric learning, the proposed method generates subordinate clusters for each data group, which are merged after finishing K-means clustering. Result of a comparison experiment with a baseline method shows the effectiveness of the proposed method in terms of success rate and NMI (normalized mutual information)