Scholarlink

<p>Detecting anomalies in attributed graphs is a crucial yet challenging task due to the interplay between node attributes, graph structure, and contextual dependencies. We propose a novel multi-stage hybrid framework, NHADF, that integrates three core modules to enhance detection capability and interpretability: (1) Embedding smoothing to mitigate local noise and inconsistency, (2) Attention-based dual encoding to fuse structural and attribute information with semantic weighting, and (3) Community-aware deviation scoring to capture contextual outliers by modeling intra-cluster residuals. Unlike existing methods that rely on single-view representations or handcrafted substructures, NHADF leverages modular learning and residual-based scoring to offer interpretable and scalable detection. Extensive evaluations on benchmark datasets demonstrate that NHADF consistently outperforms state-of-the-art baselines. For instance, on the BlogCatalog dataset, it achieves a TPR of 0.901, FPR of 0.080, and an F1-score of 0.893, indicating strong discriminative performance and robustness.</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

A multi-stage hybrid framework for anomaly detection in attributed graphs using attention-driven representation and community-aware scoring

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月14日

查看全文
  • Wasim Khan,
  • Khan Vajid Nabilal,
  • Mohammad Ishrat,
  • Kashif Asad,
  • Faheem Ahmad,
  • Meenal Suraj Wagh

摘要

Detecting anomalies in attributed graphs is a crucial yet challenging task due to the interplay between node attributes, graph structure, and contextual dependencies. We propose a novel multi-stage hybrid framework, NHADF, that integrates three core modules to enhance detection capability and interpretability: (1) Embedding smoothing to mitigate local noise and inconsistency, (2) Attention-based dual encoding to fuse structural and attribute information with semantic weighting, and (3) Community-aware deviation scoring to capture contextual outliers by modeling intra-cluster residuals. Unlike existing methods that rely on single-view representations or handcrafted substructures, NHADF leverages modular learning and residual-based scoring to offer interpretable and scalable detection. Extensive evaluations on benchmark datasets demonstrate that NHADF consistently outperforms state-of-the-art baselines. For instance, on the BlogCatalog dataset, it achieves a TPR of 0.901, FPR of 0.080, and an F1-score of 0.893, indicating strong discriminative performance and robustness.