graph-level contrastive learning, aiming to learn the representations for
each graph by contrasting two augmented graphs, has attracted considerable
attention. Previous studies usually simply assume that a graph and its
augmented graph as a positive pair, otherwise as a negative pair.