Representation Learning for Scale-Free Networks