A Fully Distributed Clustering Algorithm Based on Random Walks

Abstract

In this paper, we present a fully distributed clustering algorithm based on random walks that works on arbitrary topologies. A cluster is composed of a set of nodes called the core that coordinates the clustering process, and of non-core nodes called ordinary nodes. A core is built through a random walk based procedure. Its neighboring nodes that do not belong to any cluster are recruited by the core as ordinary nodes into its cluster. The correctness and termination of our algorithm are proven. We also prove that when two clusters are adjacent, at least one of them has a complete core (i.e. a core with the maximum size allowed by the user). Our algorithm is not deterministic, which allows a better load balancing, since the core nodes are not determined by their ids and/or location.