Anomaly Subgraph Detection on Multiple Associated Attributed Networks.

Area of Science:

• Artificial Intelligence
• Data Science
• Graph Analytics

Background:

• Anomaly subgraph detection is crucial for AI and large datasets.
• Existing methods struggle with data lacking explicit anomalous attributes.
• Implicit anomaly subgraphs (IASs) pose a significant challenge.

Purpose of the Study:

• To propose a novel approach for detecting implicit anomaly subgraphs (IASs).
• To address the limitations of existing methods in data with sparse anomalous attributes.
• To enhance the robustness and applicability of anomaly detection in complex graphs.

Main Methods:

• Utilizes transfer learning techniques to fuse features from multiple graphs.
• Employs a graph attention (GAT) network for anomaly feature extraction.
• Constructs a two-layer graph with a source graph for easier anomaly identification.

Main Results:

• Demonstrates the effectiveness and robustness of the IASD approach.
• Successfully applied to four practical anomaly subgraph detection tasks.
• Validated through experiments on five real-world datasets.

Conclusions:

• The proposed IASD method with multidimensional feature transfer is effective for detecting implicit anomalies.
• This approach overcomes limitations of traditional methods in attribute-scarce environments.
• Offers a promising solution for various real-world anomaly detection challenges.