Clustering from general pairwise observations with applications to time-varying graphs

Abstract

We present a general framework for graph clustering and bi-clustering where we are given a general observation (called a label) between each pair of nodes. This framework allows a rich encoding of various types of pairwise interactions between nodes. We propose a new tractable and robust approach to this problem based on convex optimization and maximum likelihood estimators. We analyze our algorithms under a general statistical model extending the planted partition and stochastic block models. Both sufficient and necessary conditions are provided for successful recovery of the underlying clusters. Our theoretical results subsume many existing graph clustering results for a wide range of settings, including planted partition, weighted clustering, submatrix localization and partially observed graphs. Furthermore, our results are applicable to novel settings including time-varying graphs, providing new insights to solutions of these problems. We provide empirical results on both synthetic and real data that corroborate with our theoretical findings.