Scholarlink

<p>Graph clustering aims to partition nodes into semantically coherent groups by jointly exploiting relational structure and node attribute semantics. Contrastive learning has emerged as a powerful paradigm for this task, yet its effectiveness is fundamentally constrained by two tightly coupled limitations. Prevailing methods rely on message-passing encoders with limited receptive fields, preventing them from capturing the long-range structural dependencies essential for cluster-level coherence. A natural remedy is to expand the receptive field via deeper stacking or global attention–yet this directly exacerbates a second, compounding problem: real-world graphs contain pervasive attribute noise and heterophilic edges that corrupt neighborhood signals, and broader aggregation amplifies rather than mitigates such corruption, leaving learned representations increasingly entangled across cluster boundaries. To address these synergistic challenges, we propose DiffGCC, a generative graph contrastive clustering framework that couples global–local feature encoding with a latent-space diffusion denoising mechanism. DiffGCC first removes heterophilic edges via attribute-guided topological denoising to purify the graph structure. A dual-branch encoder then fuses global long-range dependency modeling with local fine-grained aggregation (GAT) on the cleaned graph, ensuring representations are simultaneously globally informed and locally precise. A latent-space diffusion denoising module further simulates forward Gaussian noise injection and learns to reconstruct clean embeddings, providing more robust contrastive supervision and explicitly regularizing the latent space against residual noise. The framework is optimized end-to-end via a joint objective combining topology-aware similarity learning, heterogeneous view alignment, and generative denoising loss. Experiments on eight benchmarks demonstrate that DiffGCC substantially outperforms existing methods across ACC, NMI, ARI, and F1, with particularly strong gains on denser, noisier product graphs. The source code is available at: <a href="https://github.com/CisunLL/DiffGCC">https://github.com/CisunLL/DiffGCC</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DiffGCC: diffusion-enhanced global–local graph contrastive clustering

  • 期刊论文

  • 发表日期：

    2026年7月15日

查看全文
  • Lun Liu,
  • Chengyun Song

摘要

Graph clustering aims to partition nodes into semantically coherent groups by jointly exploiting relational structure and node attribute semantics. Contrastive learning has emerged as a powerful paradigm for this task, yet its effectiveness is fundamentally constrained by two tightly coupled limitations. Prevailing methods rely on message-passing encoders with limited receptive fields, preventing them from capturing the long-range structural dependencies essential for cluster-level coherence. A natural remedy is to expand the receptive field via deeper stacking or global attention–yet this directly exacerbates a second, compounding problem: real-world graphs contain pervasive attribute noise and heterophilic edges that corrupt neighborhood signals, and broader aggregation amplifies rather than mitigates such corruption, leaving learned representations increasingly entangled across cluster boundaries. To address these synergistic challenges, we propose DiffGCC, a generative graph contrastive clustering framework that couples global–local feature encoding with a latent-space diffusion denoising mechanism. DiffGCC first removes heterophilic edges via attribute-guided topological denoising to purify the graph structure. A dual-branch encoder then fuses global long-range dependency modeling with local fine-grained aggregation (GAT) on the cleaned graph, ensuring representations are simultaneously globally informed and locally precise. A latent-space diffusion denoising module further simulates forward Gaussian noise injection and learns to reconstruct clean embeddings, providing more robust contrastive supervision and explicitly regularizing the latent space against residual noise. The framework is optimized end-to-end via a joint objective combining topology-aware similarity learning, heterogeneous view alignment, and generative denoising loss. Experiments on eight benchmarks demonstrate that DiffGCC substantially outperforms existing methods across ACC, NMI, ARI, and F1, with particularly strong gains on denser, noisier product graphs. The source code is available at: https://github.com/CisunLL/DiffGCC