A Convex Formulation for Learning Scale-Free Networks via Submodular Relaxation