Faithful Interpretation for Graph Neural Networks