Scholarlink

<p>Community detection in attributed networks, the core objective is to identify groups of nodes that not only exhibit strong internal connectivity but also share similar attribute features-crucial for understanding complex systems. Most Graph Neural Network (GNN)-based methods focus on a single view, implicitly fusing structural and attribute information, making it difficult to quantify the specific contribution of each modality to node representations, and they can only capture long-range dependencies by stacking multiple GNN layers and often perform poorly when handling nodes that are highly similar in attributes but not directly connected topologically. To overcome these limitations, this paper proposes a novel Dual Graph Learning Framework for Community Detection (DGLFCD). Our core innovation is a principled dual-graph auto-encoder architecture that explicitly models and reconciles the two information modalities. Specifically, DGLFCD employs a Graph Attention Network-based auto-encoder to capture explicit topology and higher-order neighborhood information from the original graph, while Graph Convolutional Network-based auto-encoder learns from an attribute-similarity K-Nearest Neighbor (KNN) graph, which transforms isolated attribute values into a clearer relational view, reflecting the intrinsic affinities of nodes in terms of semantics, functionality, or behavior. This design aims to identify nodes that are not directly connected in the topology but exhibit high attribute similarity, capturing potential community relationships that traditional structure-based methods fail to detect. Subsequently, a dedicated fusion module aggregates the representations from the two views into a unified latent space, further enhancing the coordination and complementarity of cross-view information. Furthermore, an end-to-end self-training clustering module is incorporated to jointly optimize node embeddings and community assignments, ensuring the learned representations are explicitly geared towards the final detection task. Extensive experimental results on multiple public benchmark datasets demonstrate that DGLFCD significantly outperforms existing methods across a variety of commonly used evaluation metrics. Further ablation studies and visualization analyses verify the necessity and effectiveness of each key component. Overall, these results convincingly show that DGLFCD exhibits strong robustness, effectiveness, and generalization capability for community detection in attributed networks.</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

From nodes to communities: a novel dual graph learning framework for community detection in attributed networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月26日

查看全文
  • Yunhui Wang,
  • Mugang Lin,
  • Kan Zhang

摘要

Community detection in attributed networks, the core objective is to identify groups of nodes that not only exhibit strong internal connectivity but also share similar attribute features-crucial for understanding complex systems. Most Graph Neural Network (GNN)-based methods focus on a single view, implicitly fusing structural and attribute information, making it difficult to quantify the specific contribution of each modality to node representations, and they can only capture long-range dependencies by stacking multiple GNN layers and often perform poorly when handling nodes that are highly similar in attributes but not directly connected topologically. To overcome these limitations, this paper proposes a novel Dual Graph Learning Framework for Community Detection (DGLFCD). Our core innovation is a principled dual-graph auto-encoder architecture that explicitly models and reconciles the two information modalities. Specifically, DGLFCD employs a Graph Attention Network-based auto-encoder to capture explicit topology and higher-order neighborhood information from the original graph, while Graph Convolutional Network-based auto-encoder learns from an attribute-similarity K-Nearest Neighbor (KNN) graph, which transforms isolated attribute values into a clearer relational view, reflecting the intrinsic affinities of nodes in terms of semantics, functionality, or behavior. This design aims to identify nodes that are not directly connected in the topology but exhibit high attribute similarity, capturing potential community relationships that traditional structure-based methods fail to detect. Subsequently, a dedicated fusion module aggregates the representations from the two views into a unified latent space, further enhancing the coordination and complementarity of cross-view information. Furthermore, an end-to-end self-training clustering module is incorporated to jointly optimize node embeddings and community assignments, ensuring the learned representations are explicitly geared towards the final detection task. Extensive experimental results on multiple public benchmark datasets demonstrate that DGLFCD significantly outperforms existing methods across a variety of commonly used evaluation metrics. Further ablation studies and visualization analyses verify the necessity and effectiveness of each key component. Overall, these results convincingly show that DGLFCD exhibits strong robustness, effectiveness, and generalization capability for community detection in attributed networks.