Graph Contrastive Learning (GCL) is an emerging self-supervised learning technique on graphs, which has attracted attention for its non-dependence on costly supervised signals as well as for its strong performance in representing data features. The performance of GCL relies heavily on augmentation schemes to capture invariant patterns that remain stable under graph perturbations. However, issues towards “how to augment to ensure capturing of critical and causal patterns” remain limited. To overcome this challenge, based on causal invariant principal, we propose a hierarchical invariant graph contrastive learning framework (HI-GCL), specifically, for edge augmentation, a spectral invariant augmentation that keeps the eigenvalues of the graph’s Laplace matrix invariant, and for node augmentation, a heterogeneous augmentation scheme based on the node importance scorer and the learning of heterogeneous graph representations are developed. Furthermore, we develop a hierarchical augmentation strategy that hierarchically maintains the consistency of feature to achieve a combination of feature augmentation at the node level and edge level of the data. By comparing with the state-of-the-art 7 methods on 7 benchmark datasets, HI-GCL demonstrates its effectiveness in terms of model performance and robustness in graph classification tasks.