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

• Published in: Sensors (Basel, Switzerland) Switzerland), 2024-04, Vol.24 (8), p.2591
• Main Authors: Tan, Xiao, Yang, Jianfeng, Zhao, Zhengang, Xiao, Jinsheng, Li, Chengwang
• Format: Article
• Language: English
• Subjects: Algorithms; Behavior; Classification; Datasets; Fraud prevention; graph anomaly detection; graph convolutional neural networks; Graphs; label propagation; Machine learning; Methods; Neural networks; Propagation; semi-supervised learning; Social aspects
• ISSN: 1424-8220; 1424-8220
• DOI: 10.3390/s24082591

The era of Industry 4.0 is gradually transforming our society into a data-driven one, which can help us uncover valuable information from accumulated data, thereby improving the level of social governance. The detection of anomalies, is crucial for maintaining societal trust and fairness, yet it poses significant challenges due to the ubiquity of anomalies and the difficulty in identifying them accurately. This paper aims to enhance the performance of the current Graph Convolutional Network (GCN)-based Graph Anomaly Detection (GAD) algorithm on datasets with extremely low proportions of anomalous labels. This goal is achieved through modifying the GCN network structure and conducting feature extraction, thus fully utilizing three types of information in the graph: node label information, node feature information, and edge information. Firstly, we theoretically demonstrate the relationship between label propagation and feature convolution, indicating that the Label Propagation Algorithm (LPA) can serve as a regularization penalty term for GCN, aiding in training and enabling learnable edge weights, providing a basis for incorporating node label information into GCN networks. Secondly, we introduce a method to aggregate node and edge features, thereby incorporating edge information into GCN networks. Finally, we design different GCN trainable weights for node features and co-embedding features. This design allows different features to be projected into different spaces, greatly enhancing model expressiveness. Experimental results on the DGraph dataset demonstrate superior AUC performance compared to baseline models, highlighting the feasibility and efficacy of the proposed approach in addressing GAD tasks in the scene with extremely low proportions of anomalous data.