Scholarlink

<p>Attributed networks anomaly detection is pivotal for applications such as fraud detection, network intrusion, and spam detection. While existing methods excel in identifying attribute-based anomalies, detecting topology-based anomalies, especially in dense substructures, and addressing challenges such as class imbalance and anomaly scarcity remain significant hurdles. This paper introduces a novel framework integrating a graph attention network (GAT)-based node aggregation module, a substructure-aware topology anomaly detection module, and a data augmentation module to address these challenges. The GAT-based module dynamically adjusts the influence of neighboring nodes, capturing subtle structural deviations, while the substructure-aware module employs <i>k</i>-core decomposition to identify high-density substructures and computes anomaly scores based on intra-substructure similarity deviations. To enhance robustness and mitigate class imbalance, the data augmentation module generates synthetic node representations, enriching the training dataset and improving the detection of subtle anomalies. The refined anomaly scores, integrating insights from both topology-based and augmented predictions, are normalized for consistent interpretation across diverse datasets. The framework’s effectiveness is evaluated using AUC and AP scores on real-world datasets, demonstrating superior performance in detecting individual and collective anomalies compared to state-of-the-art methods. This approach offers a scalable, robust, and interpretable solution to anomaly 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

A robust framework for topology-based anomaly detection in attributed networks using graph attention networks, substructure analysis, and data augmentation

  • 期刊论文

  • 发表日期：

    2025年6月23日

查看全文
  • Wasim Khan,
  • Mohammad Ishrat,
  • Syed Mohd Faisal,
  • Faheem Ahmad,
  • Khan Vajid Nabilal,
  • Yogesh Kumar Sharma

摘要

Attributed networks anomaly detection is pivotal for applications such as fraud detection, network intrusion, and spam detection. While existing methods excel in identifying attribute-based anomalies, detecting topology-based anomalies, especially in dense substructures, and addressing challenges such as class imbalance and anomaly scarcity remain significant hurdles. This paper introduces a novel framework integrating a graph attention network (GAT)-based node aggregation module, a substructure-aware topology anomaly detection module, and a data augmentation module to address these challenges. The GAT-based module dynamically adjusts the influence of neighboring nodes, capturing subtle structural deviations, while the substructure-aware module employs k-core decomposition to identify high-density substructures and computes anomaly scores based on intra-substructure similarity deviations. To enhance robustness and mitigate class imbalance, the data augmentation module generates synthetic node representations, enriching the training dataset and improving the detection of subtle anomalies. The refined anomaly scores, integrating insights from both topology-based and augmented predictions, are normalized for consistent interpretation across diverse datasets. The framework’s effectiveness is evaluated using AUC and AP scores on real-world datasets, demonstrating superior performance in detecting individual and collective anomalies compared to state-of-the-art methods. This approach offers a scalable, robust, and interpretable solution to anomaly detection in attributed networks.