Graph meta-learning aims to extract cross-task transferable knowledge from meta-training tasks and improve the generalization capability on few-shot tasks. However, existing methods suffer from the heavy reliance on labeled data for meta task construction and limited graph data representation capability. To address these issues, we propose a both label-efficient and effective model, CSG-Meta (The source code is available on https://github.com/hningbo/CSG-Meta note that the footnote has been set in the following sentence “To address these issues, we propose ...”, as footnotes are not allowed in abstracts. Specifically, we first design a contrastive self-supervised meta tasks construction method to generate effective meta-training tasks without any labeled data. Then, we introduce a task-adaptive graph few-shot classifier to mitigate overfitting caused by limited labeled data and task divergence, which performs parameter initialization and transformation with prototype and task presentation. To reduce the noise in constructed tasks, we design a structural denoising module to measure the node’s confidence score with structural similarity and node importance, leading to more accurate prototype and task representation. Experimental results on four real-world attributed graph datasets show that CSG-Meta achieves better node classification performance on most few-shot learning scenarios. On the Amazon-Clothing dataset, the improvement rates of accuracy are from 4.4% to 9.6% compared to the current state-of-the-art supervised graph meta-learning model COSMIC.