Find Your Place: Simple Distributed Algorithms for Community Detection

Given an underlying network, the {\em averaging dynamics} is the following distributed process: Initially, each node sets its local state to an element of $\{-1,1\}$, uniformly at random and independently of other nodes. Then, in each consecutive round, every node updates its state to the average of its neighbors. We show that when the graph is organized into two equal-sized "communities" separated by a sparse cut and certain additional conditions hold, a global state is reached within a logarithmic number of rounds, in which states of nodes belonging to the same community lie within a small range, while states of members of different communities are separated by a large gap. Depending on our assumptions on the underlying graph, this can either hold for all or for most nodes. Either way, this behavior allows nodes to locally identify (exactly or approximately) the structure of the two communities. We show that the conditions under which the above happens are satisfied with high probability in graphs sampled according to the stochastic block model $\mathcal{G}_{n,p,q}$, provided that $\min \{ (p - q)n, qn \} \gg \sqrt { (p+q)n \log n}$, in which case the nodes are able to perform approximate reconstruction of the community structure in $\mathcal{O}(\log n)$ time. We further prove that, with respect to an analog of the stochastic block model for regular graphs, an algorithm based on the averaging dynamics performs exact reconstruction in a logarithmic number of rounds for the full range of parameters for which exact reconstruction is known to be doable in polynomial time.