BLISS: Bandit Layer Importance Sampling Strategy for Efficient Training of Graph Neural Networks