Distributed Submodular Maximization: Identifying Representative Elements in Massive Data