Towards distributed ensemble clustering for networked sensing systems

Given a set of different clustering solutions to a unified dataset, ensemble clustering is to aggregate them to yield a more accurate and robust solution. In recent years, ensemble clustering has been extensively studied and successfully applied to many areas. In this paper, we study a new variant of ensemble clustering, distributed ensemble clustering, motivated by the proliferation of networked sensing systems where communication is enabled between only connected nodes. Our goal is to aggregate the clustering solutions produced by the sensor nodes that observe the same set of objects. Different from traditional ensemble clustering problems, distributed ensemble clustering aims to achieve not only accurate clustering results, but also low communication cost among the nodes. To this end, we build a novel geometric optimization model that can be efficiently solved with theoretical quality guarantee. The proposed approach, bearing nice geometric properties, can be easily adapted to distributed settings without any sacrifice of clustering quality, and facilitates a dimension reduction procedure which can significantly reduce the communication complexity. We validate our approach on two benchmark datasets. Experimental results suggest that our approach can efficiently solve the distributed ensemble clustering problem, and outperform the baselines on both clustering accuracy and communication cost.