Основные понятия
Graph anomaly detection methods are enhanced by novel outlier injection techniques, message passing analysis, and the utilization of hyperbolic neural networks.
Аннотация
The content discusses outlier injection methods, message passing impact, and hyperbolic neural networks in node-level graph anomaly detection. It covers the challenges, strategies, and experimental results in detail.
Three Revisits to Node-Level Graph Anomaly Detection:
- Introduction to the importance of graph anomaly detection.
- Challenges in benchmarking approaches for unsupervised node-level graph anomaly detection.
- Revisiting datasets and approaches for unsupervised node-level graph anomaly detection tasks.
Outlier Injection Methods:
- Review of previous outlier injection methods and their limitations.
- Introduction of new outlier injection methods based on graph information.
- Comparison of outlier detection using norm information for different outlier injection methods.
Strategies for Outlier Detection:
- Analysis of the impact of message passing on outlier detection.
- Introduction of hyperbolic neural networks for preserving geometry in outlier detection.
- Description of feature transformation, centralization, losses, and training in hyperbolic models.
Статистика
"The optimal MLPAE has zero reconstruction error for normal nodes and unit reconstruction error for anomalous nodes."
"The optimal GCNAE has reconstruction error √2(1−nnormal/nV) for normal nodes and √2nnormal/nV reconstruction error for anomalous nodes."
Цитаты
"Our study sheds light on general strategies for improving node-level graph anomaly detection methods."
"The negative impact of message passing on performance is further evidenced across other scenarios."