Scholarlink

<p>Deep attributed graph clustering aims to group nodes into semantically meaningful clusters by leveraging both node attributes and complex topological structures. However, most existing methods rely on static adjacency matrices and are limited by the local aggregation scope of graph neural networks. This leads to noisy structural features that fail to capture global cluster structures, which in turn degrades clustering performance. To address these challenges, we propose a dual-driven network that enhances the robustness and discriminability of graph clustering through dynamic structure optimization and prototype contrastive learning. The model adopts a co-training dual-branch framework that integrates attribute representations from an autoencoder with structural features from a graph autoencoder, achieving unified optimization from low-level features to high-order topology. Meanwhile, the graph structure is iteratively updated based on node features and current topology, enabling adaptive reconstruction and noise suppression. A momentum-based prototype contrastive module is further introduced to continuously align cluster prototypes in the latent space, improving intra-cluster compactness and inter-cluster separability. The entire model is trained under a unified self-supervised objective. Finally, traditional clustering algorithms are applied to the learned embeddings for node partitioning. Experiments on five benchmark datasets show that our method consistently outperforms 13 state-of-the-art clustering approaches. In particular, it achieves a 2% clustering accuracy improvement over the best existing method on the DBLP dataset. Our code is available at <a href="https://github.com/aaaaaaa1wqdqw/DynStruct-Cluster">https://github.com/aaaaaaa1wqdqw/DynStruct-Cluster</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

Attribute graph clustering via dual-driven network with dynamic structure optimization and prototype contrast

  • 期刊论文

  • 发表日期：

    2025年10月25日

查看全文
  • Shuqiu Tan,
  • Huibing Li

摘要

Deep attributed graph clustering aims to group nodes into semantically meaningful clusters by leveraging both node attributes and complex topological structures. However, most existing methods rely on static adjacency matrices and are limited by the local aggregation scope of graph neural networks. This leads to noisy structural features that fail to capture global cluster structures, which in turn degrades clustering performance. To address these challenges, we propose a dual-driven network that enhances the robustness and discriminability of graph clustering through dynamic structure optimization and prototype contrastive learning. The model adopts a co-training dual-branch framework that integrates attribute representations from an autoencoder with structural features from a graph autoencoder, achieving unified optimization from low-level features to high-order topology. Meanwhile, the graph structure is iteratively updated based on node features and current topology, enabling adaptive reconstruction and noise suppression. A momentum-based prototype contrastive module is further introduced to continuously align cluster prototypes in the latent space, improving intra-cluster compactness and inter-cluster separability. The entire model is trained under a unified self-supervised objective. Finally, traditional clustering algorithms are applied to the learned embeddings for node partitioning. Experiments on five benchmark datasets show that our method consistently outperforms 13 state-of-the-art clustering approaches. In particular, it achieves a 2% clustering accuracy improvement over the best existing method on the DBLP dataset. Our code is available at https://github.com/aaaaaaa1wqdqw/DynStruct-Cluster.