Measuring and maximizing group closeness centrality over disk-resident graphs

As an important metric in graphs, group closeness centrality measures how close a group of vertices is to all other vertices in a graph, and it is used in numerous graph applications such as measuring the dominance and influence of a node group over the graph. However, when a large-scale graph contains hundreds of millions of nodes/edges which cannot reside entirely in computer's main memory, measuring and maximizing group closeness become challenging tasks. In this paper, we present a systematic solution for efficiently calculating and maximizing the group closeness for disk- resident graphs. Our solution first leverages a \probabilistic counting method" to efficiently estimate the group closeness with high accuracy, rather than exhaustively computing it in an exact fashion. In addition, we design an I/O-efficient greedy algorithm to find a node group that maximizes group closeness. Our proposed algorithm significantly reduces the number of random accesses to disk, thereby dramatically im- proving computational efficiency. Experiments on real-world big graphs demonstrate the efficacy of our approach.