Learning Sparse Combinatorial Representations via Two-stage Submodular Maximization