Dual-level noise augmentation for graph clustering with triplet-wise contrastive learning

Contrastive deep graph clustering has attracted widespread attention due to its self-supervised representation learning mechanism and excellent clustering performance. Although, most existing methods rely on low-pass filtering to achieve denoising, ignoring the potential benefit of high-frequency information and noise in obtaining comprehensive and robust representation. Second, commonly used contrastive learning strategies generally treat non-target samples as negative samples, which is prone to triggering contrastive bias and weakening the representation quality. To this end, this paper proposes a novel contrastive graph clustering framework, Dual-level Noise Augmentation for Graph Clustering with triplet-wise Contrastive learning (DNA-CGC), strengthening the benefit of noise to enrich the representation learning and amplify the contrastive learning efficacy. It consists of two core modules, Hybrid Noise Representation Augmentation (HNRA) and Noise-Aware Contrastive Learning (NACL). The HNRA module integrates low- and high-frequency graph signals to capture both shared and distinctive node characteristics, while introducing Gaussian noise as beneficial perturbation to enrich the representation diversity, thereby achieving multi-information fusion under hybrid noise. The NACL module, on the other hand, generates exclusive negative samples through Gaussian noise and constructs the triplet-wise contrastive pairs (Target, Positive, Negative), mitigating the contrastive bias by preventing false negatives and further facilitating more accurate semantic alignment. Extensive experiments on six benchmark datasets validate the significant advantages of DNA-CGC in terms of clustering performance and representation quality. The code could be available at https://github.com/TianxiangZhao0474/DNA-CGC.git.