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

Abstract

Unsupervised graph-level anomaly detection (UGLAD) has garnered increasing attention in recent years due to its practical significance. However, most existing methods primarily rely on GNNs that consider only pairwise relationships between first-order neighbors. This is insufficient to capture complex high-order dependencies often associated with anomalies. Furthermore, these methods are confined to Euclidean space and ignore common underlying properties (e.g., hierarchies) in real-world datasets, easily leading to high-distortion representations and compromising detection performance. To address these limitations, we propose a novel Dual Hyperbolic Contrastive Learning Framework for Unsupervised Graph-Level Anomaly Detection (HC-GLAD in short). We introduce a new anomaly-capturing perspective from high-order group information via hypergraphs to identify subtle anomalies in a complementary manner. Furthermore, we introduce hyperbolic geometry and design hyperbolic contrastive learning within a dual architecture for UGLAD, fully exploiting the capacity advantage of hyperbolic space and considering underlying hierarchies in graphs for more discriminative and low-distortion representations. To the best of our knowledge, this is the first work to consider node group information and hierarchies for UGLAD, simultaneously introducing hypergraph and hyperbolic learning to the UGLAD task. Extensive experiments on nine real-world datasets from diverse domains demonstrate the superiority of HC-GLAD. The code is available at https://github.com/Yali-Fu/HC-GLAD.