Detecting dense communities in large social and information networks with the Core & Peel algorithm

Detecting and characterizing dense subgraphs (tight communities) in social and information networks is an important exploratory tool in social network analysis. Several approaches have been proposed that either (i) partition the whole network into clusters, even in low density region, or (ii) are aimed at finding a single densest community (and need to be iterated to find the next one). As social networks grow larger both approaches (i) and (ii) result in algorithms too slow to be practical, in particular when speed in analyzing the data is required. In this paper we propose an approach that aims at balancing efficiency of computation and expressiveness and manageability of the output community representation. We define the notion of a partial dense cover (PDC) of a graph. Intuitively a PDC of a graph is a collection of sets of nodes that (a) each set forms a disjoint dense induced subgraphs and (b) its removal leaves the residual graph without dense regions. Exact computation of PDC is an NP-complete problem, thus, we propose an efficient heuristic algorithms for computing a PDC which we christen Core and Peel. Moreover we propose a novel benchmarking technique that allows us to evaluate algorithms for computing PDC using the classical IR concepts of precision and recall even without a golden standard. Tests on 25 social and technological networks from the Stanford Large Network Dataset Collection confirm that Core and Peel is efficient and attains very high precison and recall.