Efficient evolution-based clustering of high dimensional data streams with dimension projection

Abstract

SE-Stream is an evolution-based stream clustering method that supports high dimensional data streams. SE-Stream is able to monitor and detect change in the clustering structure during the progression of data streams. In this paper, we improve performance of SE-Stream by reducing its execution time and increasing its cluster quality. SE-Stream reduces complexity of stream processing by determining appropriated subset of dimensions of each active cluster to express cluster specific characteristics during the progression of data streams. With elimination of redundant operations, SE-Stream is improved both in terms of cluster quality and execution time. Experimental results on two real-world datasets show that SE-Stream outperforms its previous version in terms execution time. Further, the cluster quality in terms of both purity and f-measure has been considerably improved. Compared with HPStream, a state of the art algorithm for projected clustering of high dimensional data streams, SE-Stream outperforms in terms of cluster quality and yields comparable execution time.