A new online clustering approach for data in arbitrary shaped clusters

Abstract

In this paper we demonstrate a new density based clustering technique, CODSAS, for online clustering of streaming data into arbitrary shaped clusters. CODAS is a two stage process using a simple local density to initiate micro-clusters which are then combined into clusters. Memory efficiency is gained by not storing or re-using any data. Computational efficiency is gained by using hyper-spherical micro-clusters to achieve a micro-cluster joining technique that is dimensionally independent for speed. The micro-clusters divide the data space in to sub-spaces with a core region and a non-core region. Core regions which intersect define the clusters. A threshold value is used to identify outlier micro-clusters separately from small clusters of unusual data. The cluster information is fully maintained on-line. In this paper we compare CODAS with ELM, DEC, Chameleon, DBScan and Denstream and demonstrate that CODAS achieves comparable results but in a fully on-line and dimensionally scale-able manner.