<p>Self-supervised learning empowers models to learn from unlabeled data, making it a popular technique for addressing the challenge of insufficient labeled data. In this paper, we combine the issue of cold items lacking supervision signals due to the long-tail distribution in recommendation systems, which can lead to model degradation, and propose graph self-supervised long-tail item augmentation for recommendation (SLIA). We study the proxy task of distinguishing between popular items and cold items and employ various graph structure perturbation schemes on the user–item interaction graph (U–I graph) for contrastive learning. To be more specific, we propose a proportional edge dropping strategy for popular items. For cold items, while retaining their sparse interaction data, we introduce random perturbations to create a new enhanced U–I graph. Through the fine-grained construction of contrastive views, this scheme ensures a global perturbation of the original views and prevents the loss of information for cold items in the tail of the power-law distribution. Additionally, we propose a contrastive loss with a punishment mechanism to explicitly enhance the invariance of node representations. Through theoretical analysis, we find that SLIA is not only effective in mining hard negative samples but also reduces the occurrence of outliers in the representation distribution space. Experimental results on four benchmark datasets demonstrate that SLIA achieves competitive results in both supervised and self-supervised recommendation algorithms.</p>