Do We Really Need Contrastive Learning for Graph Representation?