Evaluating the stability of communities found by clustering algorithms

Abstract

Since clustering algorithms identify communities even in networks that are not believed to exhibit clustering behavior, we are interested in evaluating the significance of the communities returned by these algorithms. As a proxy for significance, prior work has investigated stability: by how much do the clusters change when the network is perturbedWe describe a cheap and simple method for evaluating stability and test it on a variety of real-world and synthetic graphs. Rather than evaluating our method on a single clustering algorithm, we give results using three popular clustering algorithms and measuring the distance between computed clusterings via three different metrics. We consider the results in the context of known strengths and weaknesses of the three clustering algorithms and also compare our method against that of measuring modularity. Finally we give evidence that our method provides more information than does the modularity. © Springer-Verlag Berlin Heidelberg 2013.