DAMR: Multi-scale graph contrastive learning with dynamic adjustment and mutual rectification

Graph contrastive learning (GCL) has emerged as a powerful self-supervised approach for learning generalized graph representations, achieving remarkable advancements in recent years. However, most existing GCL methods ignore the noise of the augmented global structure and the dynamic change in training, and lack detailed consideration in calculating local structural homogeneity. These limitations may lead to the model’s insufficient performance in capturing fine-grained semantic features at the node level, making it difficult to fully explore the potential semantic associations between adjacent nodes. Meanwhile, on a global scale, there is also a lack of the ability to model complex topological structures. To this end, we propose a new multi-scale graph contrastive learning with dynamic adjustment and mutual rectification. This method dynamically adjusts the global structure via graph reconstruction and adaptively learns node representations; Meanwhile, a mutual rectification module is designed to predict the support scores of neighbors relative to anchors and quantify each neighbor’s contribution to view agreement. Both reconstruction and rectification are integrated into the training objective and effectively capture the graph structure information from both global and local scales, improving the quality and robustness of graph representations. We conduct extensive experiments on three downstream tasks: node classification, node clustering, and link prediction. The experimental results demonstrate that our method outperforms existing GCL methods across multiple tasks and datasets, validating the effectiveness and generalizability of the proposed model.