Improving Graph Convolutional Network with Learnable Edge Weights and Edge-Node Co-Embedding for Graph Anomaly...

Area of Science:

• Artificial Intelligence
• Data Science
• Network Analysis

Background:

• Industry 4.0 generates vast data, necessitating robust anomaly detection for social governance and trust.
• Existing Graph Anomaly Detection (GAD) methods struggle with datasets containing extremely low proportions of anomalous labels.
• Accurate anomaly identification is challenging due to their ubiquity and the difficulty in precise detection.

Purpose of the Study:

• To enhance the performance of Graph Convolutional Network (GCN)-based GAD algorithms on datasets with scarce anomalous labels.
• To fully leverage node label, node feature, and edge information within graph structures for improved anomaly detection.
• To develop a more expressive and effective GAD model for data-driven societal governance.

Main Methods:

• Modified GCN network structure and feature extraction techniques were employed.
• The relationship between Label Propagation Algorithm (LPA) and feature convolution was theoretically established, enabling LPA as a GCN regularization term.
• A method for aggregating node and edge features was introduced, alongside distinct GCN trainable weights for node and co-embedding features to enhance model expressiveness.

Main Results:

• The proposed approach demonstrated superior Area Under the Curve (AUC) performance on the DGraph dataset compared to baseline models.
• The integration of label propagation and feature aggregation significantly improved anomaly detection capabilities.
• Experimental results validated the feasibility and efficacy of the modified GCN for GAD in low-proportion anomaly scenarios.

Conclusions:

• The developed GCN-based GAD method effectively addresses the challenge of detecting anomalies in datasets with extremely low anomalous label proportions.
• The approach successfully integrates diverse graph information (node labels, features, edges) for enhanced detection accuracy.
• This work provides a valuable tool for improving social governance and maintaining trust in the data-driven era of Industry 4.0.