The Dynamic Hyper-ellipsoidal Micro-Clustering for Evolving Data Stream Using Only Incoming Datum

Abstract

Data stream clustering is becoming the efficient method to cluster an online massive data. The clustering task requires a process capable of partitioning data continuously with incremental learning method. In this paper, we present a new clustering method, called DyHEMstream, which is online and offline algorithm. In online phase, dynamic hyper-ellipsoidal micro-cluster is proposed used to keep summary information about evolving data stream based on new incoming data sample. The shape of proposed micro-cluster can represent the incoming data better than traditional micro-cluster. The algorithm processes each data point in one-pass fashion without storing the entire data set. In offline phase, each cluster is generated by expanding hyper-ellipsoidal micro-clusters to form the final clusters. The DyHEMstream algorithm is evaluated on various synthetic data sets using different quality metrics compared with a famous data stream clustering -- DenStream. Based on purity, Rand index, and Jaccard index, DyHEMstrem is very efficient than DenStream in term of clustering quality in different shapes, sizes, and densities in noisy data.