Uncovering Capabilities of Model Pruning in Graph Contrastive Learning