ANOMIX: A Simple yet Effective Hard Negative Generation via Mixing for Graph Anomaly Detection

TL;DR

This work proposes ANOMIX, a framework that consists of a novel graph mixing approach, ANOMIX-M, and multi-level contrasts for GAD, which can effectively mix abnormality and normality from input graph to generate hard negatives, which are important for efficient GCL.

Abstract

Graph contrastive learning (GCL) generally requires a large number of samples. The one of the effective ways to reduce the number of samples is using hard negatives (e.g., Mixup). Designing mixing-based approach for GAD can be difficult due to imbalanced data or limited number of anomalies. We propose ANOMIX, a framework that consists of a novel graph mixing approach, ANOMIX-M, and multi-level contrasts for GAD. ANOMIX-M can effectively mix abnormality and normality from input graph to generate hard negatives, which are important for efficient GCL. ANOMIX is (a) A first mixing approach: firstly attempting graph mixing to generate hard negatives for GAD task and node- and subgraph-level contrasts to distinguish underlying anomalies. (b) Accurate: winning the highest AUC, up to 5.49% higher and 1.76% faster. (c) Effective: reducing the number of samples nearly 80% in GCL. Code is available at https://github.com/missinghwan/ANOMIX.

Figures (6)

• Figure 1: An example of generating hard negatives for image (up) and graph (down).
• Figure 2: An overview of A NOM IX.
• Figure 3: ROC curves on four datasets.
• Figure 4: Effect of balance parameters, contamination ratio, and hardness level.
• Figure 5: Effect of our hard negatives on ACM dataset.