Community Detection on Euclidean Random Graphs

In this paper, Abishek Sankararaman and Fran├žois Baccelli introduce the problem of Community Detection on a new class of sparse spatial random graphs embedded in the Euclidean space. They consider the planted partition version of the random connection model graph studied in classical stochastic geometry. The nodes of the graph form a marked Poisson Point Process of intensity \lambda with each node being equipped with an i.i.d. uniform mark drawn from {-1,+1}. Conditional on the labels, edges are drawn independently at random depending both on the Euclidean distance between the nodes and the community labels on the nodes. The paper studies the Community Detection problem on this random graph which consists in estimating the partition of nodes into communities, based on an observation of the random graph along with the spatial location labels on nodes. For all dimensions greater than or equal to 2, they establish a phase transition in the intensity of the point process. They show that if the intensity is small, then no algorithm for Community Detection can beat a random guess. This is proven by introducing and analyzing a new problem called ‘Information Flow from Infinity’. On the positive side, they give an efficient algorithm to perform Community Detection as long as the intensity is sufficiently high. Along the way, a distinguishability result is established, which says that one can always infer the presence of a partition, even when one cannot identify the partition better than at random. This is a surprising new phenomenon not observed thus far in any non spatial Erdos-Renyi based planted partition models.