Real-positive-neighbors guide Contrastive Graph Clustering Network

The rapid advancement of deep learning has introduced promising techniques for attribute graph clustering. However, existing deep attributed graph clustering methods face two key limitations: (1) insufficient exploration of multi-scale neighborhood structural information during training, and (2) inappropriate graph data augmentation strategies, which often lead to semantic drift and indistinguishable positive samples. To address these issues, this paper proposes a novel Real-positive-neighbors Guided Contrastive Graph Clustering Network (ReCogNet) for attribute graph clustering. ReCogNet employs a dynamic attention-weighted fusion mechanism to refine shallow semantic information derived from the multi-scale GCN network, enabling the model to capture subtle yet critical node relationships. Additionally, it dynamically identifies real-positive-neighbor nodes and adopts a negative-free contrastive learning objective. This objective maximizes the similarity between a query node and its real-positive-neighbors in the latent embedding space, thereby improving clustering performance by leveraging meaningful local relationships. Extensive experiments on six benchmark datasets demonstrate that the proposed ReCogNet method consistently outperforms state-of-the-art approaches.