Gossip-Based Multidimensional Data Clustering and Its Applications

In this paper, two multi-dimensional data clustering methods based on gossip algorithms describing the opinion forming process over network of agent through pairwise interaction are proposed. The two methods are called Simple Clustering and Consensus Clustering. The later one takes into account the robustness of discovered clusterings derived from aggregating multiple runs from the former. The convergence of the methods are guaranteed by that of the gossip algorithm. We illustrate the application of the proposed methods on a variety of real problem: classifying the energy power demand behavior among users which is helpful for electricity demand prediction in energy management system, and classifying behavior of batteries from the recorded data, which is potential for abnormal detection.