A cognition-inspired system for data stream clustering

In applications such as target detection, domain knowledge of sensed data is often available. In this paper, we incorporate the available domain knowledge into clustering process and develop a knowledge-driven Mahalanobis distance-based ART (adaptive resonance theory) clustering algorithm. The strength of the knowledge-driven algorithm is that it can automatically determine the number of clusters with improved clustering results. The validity of the new algorithm has been verified on four artificial datasets. In addition, the algorithm has been adopted in our cognition-inspired system for clustering data stream, where known target library and dispersion of feature or attributes are available. The basic idea of this system is to divide data stream into frames, and to incorporate knowledge learned in previous frames into clustering of the following ones. Experimental studies have demonstrated that the evolving learning mechanism leads to improved clustering results compared with conventional incremental clustering algorithm Fuzzy ART and batch-based clustering algorithm k-means. © 2015 SERSC.