Scholarlink

<p>Due to the widespread application of dynamic graphs in areas such as social networks and e-commerce, research on anomaly detection in dynamic graphs has attracted increasing attention. Existing methods usually sample neighbor nodes with a fixed number of hops and do not consider long-term historical information. This paper proposes a graph representation learning approach using diffusion technique for dynamic graph anomaly detection, abbreviated as GDDyAD. The proposed method aims to detect anomalous edges in dynamic graphs, which enables more accurate subgraph sampling by assessing the relative importance of nodes, thereby reducing the impact of noise from irrelevant nodes and lowering computational costs. Simultaneously, it comprehensively integrates long-term historical information. Specifically, for each node, we selected the top-<i>k</i> most influential nodes for sampling based on diffusion technology within a time window. Subsequently, Graph Convolutional Networks (GCNs) are used to obtain necessary structural information. Moreover, Gated Recurrent Units(GRUs) are employed to capture temporal information and attention weights are assigned to each GRU to take into account the impact of historical information. Finally, pooling layers are utilized to acquire the final representations of the target nodes for anomaly detection. The GDDyAD method is evaluated on six publicly available datasets, and the results indicate that GDDyAD outperforms existing baseline methods, with an average improvement of 2.17% on AUC values. The code is available at <a href="https://github.com/wmwm6666/GDDyAD">https://github.com/wmwm6666/GDDyAD</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

Graph representation learning method using diffusion sampling for dynamic graph anomaly detection

  • 期刊论文

  • 发表日期：

    2025年10月17日

查看全文
  • Jing He,
  • Mian Wang,
  • Zhen Zhang,
  • Lin Chen,
  • YunYun Dong

摘要

Due to the widespread application of dynamic graphs in areas such as social networks and e-commerce, research on anomaly detection in dynamic graphs has attracted increasing attention. Existing methods usually sample neighbor nodes with a fixed number of hops and do not consider long-term historical information. This paper proposes a graph representation learning approach using diffusion technique for dynamic graph anomaly detection, abbreviated as GDDyAD. The proposed method aims to detect anomalous edges in dynamic graphs, which enables more accurate subgraph sampling by assessing the relative importance of nodes, thereby reducing the impact of noise from irrelevant nodes and lowering computational costs. Simultaneously, it comprehensively integrates long-term historical information. Specifically, for each node, we selected the top-k most influential nodes for sampling based on diffusion technology within a time window. Subsequently, Graph Convolutional Networks (GCNs) are used to obtain necessary structural information. Moreover, Gated Recurrent Units(GRUs) are employed to capture temporal information and attention weights are assigned to each GRU to take into account the impact of historical information. Finally, pooling layers are utilized to acquire the final representations of the target nodes for anomaly detection. The GDDyAD method is evaluated on six publicly available datasets, and the results indicate that GDDyAD outperforms existing baseline methods, with an average improvement of 2.17% on AUC values. The code is available at https://github.com/wmwm6666/GDDyAD.