Scholarlink

<p>Social networks play a crucial role in communication, information exchange, and the development of online communities. However, their growth has also intensified the challenge of identifying anomalous communities, which may indicate security threats or malicious activities. This study aims to develop an effective and interpretable framework for accurately detecting such communities across diverse network types. We introduce a four-phase self-supervised deep learning approach, where, first, a multilayer AutoEncoder extracts compact and noise-reduced feature representations. Second, a Graph Attention Network (GAT) models structural dependencies by adaptively weighting node neighborhoods. Third, a hybrid clustering strategy combining K-Means and particle swarm optimization (PSO) generates high-quality pseudo-labels for unlabeled data. Finally, an Anomaly Score Estimation (ASE) mechanism quantifies community abnormality based on cluster consistency and neighborhood separation. The proposed method was evaluated on six benchmark datasets spanning graph, image, sensor, and text modalities, and compared with two state-of-the-art baselines (DGC and GLADC). Experimental findings demonstrate consistent performance gains of up to 8% in accuracy, 6% in F1-score, 5% in NMI, and 9% in ARI, confirming its robustness across heterogeneous domains. The integration of AutoEncoder-based feature learning, GAT-driven structural modeling, optimized clustering, and anomaly scoring yields a generalizable and interpretable solution for detecting anomalous communities. The framework offers practical applicability in social media monitoring, cybersecurity, and other networked systems where timely detection is critical.</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

Anomalous community detection in social networks using graph attention and particle swarm optimization-based clustering

  • 期刊论文

  • 发表日期：

    2025年12月9日

查看全文
  • Safoura Akhlaghi

摘要

Social networks play a crucial role in communication, information exchange, and the development of online communities. However, their growth has also intensified the challenge of identifying anomalous communities, which may indicate security threats or malicious activities. This study aims to develop an effective and interpretable framework for accurately detecting such communities across diverse network types. We introduce a four-phase self-supervised deep learning approach, where, first, a multilayer AutoEncoder extracts compact and noise-reduced feature representations. Second, a Graph Attention Network (GAT) models structural dependencies by adaptively weighting node neighborhoods. Third, a hybrid clustering strategy combining K-Means and particle swarm optimization (PSO) generates high-quality pseudo-labels for unlabeled data. Finally, an Anomaly Score Estimation (ASE) mechanism quantifies community abnormality based on cluster consistency and neighborhood separation. The proposed method was evaluated on six benchmark datasets spanning graph, image, sensor, and text modalities, and compared with two state-of-the-art baselines (DGC and GLADC). Experimental findings demonstrate consistent performance gains of up to 8% in accuracy, 6% in F1-score, 5% in NMI, and 9% in ARI, confirming its robustness across heterogeneous domains. The integration of AutoEncoder-based feature learning, GAT-driven structural modeling, optimized clustering, and anomaly scoring yields a generalizable and interpretable solution for detecting anomalous communities. The framework offers practical applicability in social media monitoring, cybersecurity, and other networked systems where timely detection is critical.