HC-GLAD: Dual hyperbolic contrastive learning for unsupervised graph-level anomaly detection.

Area of Science:

• Graph Machine Learning
• Anomaly Detection
• Hypergraph Theory
• Hyperbolic Geometry

Background:

• Unsupervised graph-level anomaly detection (UGLAD) is crucial but challenged by limitations of existing methods.
• Current approaches using Graph Neural Networks (GNNs) focus on pairwise relationships, failing to capture complex high-order dependencies.
• Existing methods are confined to Euclidean space, ignoring data hierarchies and leading to high-distortion representations.

Purpose of the Study:

• To propose a novel Dual Hyperbolic Contrastive Learning Framework for Unsupervised Graph-Level Anomaly Detection (HC-GLAD).
• To address limitations in capturing high-order dependencies and data hierarchies in UGLAD.
• To develop a method that provides more discriminative and low-distortion representations for anomaly detection.

Main Methods:

• Introduced hypergraphs to capture high-order group information for anomaly identification.
• Designed a dual architecture employing hyperbolic contrastive learning within hyperbolic geometry.
• Exploited the capacity of hyperbolic space and graph hierarchies for improved representation learning.

Main Results:

• HC-GLAD demonstrates superior performance on nine real-world datasets across diverse domains.
• The framework effectively identifies subtle anomalies by incorporating node group information and hierarchies.
• Hyperbolic learning and hypergraphs contribute to more discriminative and low-distortion representations.

Conclusions:

• HC-GLAD represents a significant advancement in unsupervised graph-level anomaly detection.
• The integration of hypergraph and hyperbolic learning offers a novel and effective approach to UGLAD.
• The proposed method addresses key limitations of existing GNN-based techniques.