Learning Interpretable Decision Rule Sets: A Submodular Optimization Approach