Title: A Compressive Sensing Approach to Cluster Extraction

Abstract: We use techniques from compressive sensing to design a cluster extraction algorithm by treating the cluster indicator vector as a sparse solution to a linear system involving the graph Laplacian. We show that our algorithm, which we call SingleClusterPursuit, has computational complexity $O(d_{\max}n\log(n))$ and verify numerically that it runs quickly and accurately on large data sets, both synthetic and real. We prove that, for a large class of probabilistic models of graphs with clusters, which includes the popular stochastic block model, the fraction of vertices misclassified by SingleClusterPursuit goes to zero as the size of the cluster goes to infinity.