Giannakis, “Consensus-based k-means algorithm for distributed learning using wireless sensor networks

Clustering is a critical task in distributed exploratory and data mining scenarios whose goal is obtaining clusters of obser-vations that best reveal underlying structures. Most often, wireless sensor networks offer considerable amounts of multi-dimensional observations rendering centralized clustering ap-proaches impractical. This paper develops two decentralized k-means algorithms for clustering observations collected at spatially deployed wireless sensors. In both algorithms, sen-sors exchange sufficient information only with their one-hop neighbors. Surprisingly, simulations reveal that distributed k-means are less sensitive to initialization than their centralized counterparts resulting in superior performance with respect to the centralized k-means algorithm. Index Terms — distributed learning, wireless sensor networks 1