Dynamic Evolving Cluster Models Using On-line Split-and-Merge Operations

In this paper, we propose a new dynamic split-and-merge concept for evolving prototype-based cluster models, i.e. cluster partitions which are incrementally learned and extended on-the-fly from data streams. New criteria when clusters should be merged are based on a touching and on a homogeneity condition between two ellipsoidal clusters, the merging itself is conducted by using weighted averaging of cluster centers and a convex combination of cluster spreads based on the recursive variance update concept. The splitting criterion for an updated cluster employs a 2-means algorithm on its sub-samples and compares the quality of the split cluster with the original cluster by using Bayesian information criterion, the cluster partition with the better quality remains for the next incremental update cycle. The results on 2-dimensional as well high-dimensional streaming clustering data sets show that the new split-and-merge concept is able to produce more reliable cluster partitions than conventional evolving clustering.