Higher-order Enhanced Contrastive-based Graph Anomaly Detection Without Graph Augmentation

doi:10.1016/j.patcog.2025.111666

Higher-order Enhanced Contrastive-based Graph Anomaly Detection Without Graph Augmentation

2025-05-29

0

PRE

AI

DOI：10.1016/j.patcog.2025.111666

Original

Original request for help

Save

Abstract

En 中文

Graph anomaly detection (GAD) has been widely applied in various attributed graph data mining domains, such as financial fraud and academic citation networks. Its goal is to detect instances in graph data significantly different from other data. Existing GAD methods often modify node features or structures based on graph augmentation to improve the robustness of the model, which may lead to poor model performance and efficiency due to anomaly confusion. To tackle the aforementioned issues, we propose a novel Higher-order Enhanced Contrastive-based Graph Anomaly Detection without Graph Augmentation (HEC-GAD) in this paper. To avoid anomaly distortion, we no longer perform graph augmentation, but instead sample from ego-net and expand it to high-order neighbors to generate multi-view subgraph of the target node. To more efficiently mine anomalies in topology and attribute information, we conduct intra-view node-subgraph and inter-view subgraph-subgraph contrastive learning, and identify abnormal nodes through the reconstruction error of nodes. Comprehensive experimental results on eight artificial and real-world benchmark datasets demonstrate the effectiveness of our proposed higher-order enhanced multi-view contrastive graph anomaly detection method.

Keywords:

Graph anomaly detection

Graph contrastive learning

Multi-view contrastive learning

Higher-order Enhanced Contrastive-based Graph Anomaly Detection Without Graph Augmentation

Abstract

Journal

Organization