A Note on Over-Smoothing for Graph Neural Networks