MoEGAD: A Mixture-of-Experts Framework With Pseudo-Anomaly Generation for Graph-Level Anomaly Detection

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Graph-level anomaly detection (GLAD) identifies outlier graphs but struggles with scarce labeled anomalies.
• Limited anomaly diversity hinders robust decision boundary learning in semi-supervised GLAD.
• Multi-task graph anomaly detection remains an underexplored but crucial area.

Purpose of the Study:

• To propose MoEGAD, a novel framework for graph-level anomaly detection (GLAD).
• To address the challenges of limited labeled anomalies and enhance multi-task GLAD capabilities.
• To leverage a mixture of experts (MoE) architecture for improved GLAD performance.

Main Methods:

• An iterative anomalous graph generation module creates pseudo-anomalies for training.
• An early stopping mechanism ensures generated anomalies are sufficiently dissimilar from normal graphs.
• A latent MoE module with expert and gating networks enables cross-task adaptability.

Main Results:

• MoEGAD significantly outperforms state-of-the-art GLAD baselines in experiments.
• The framework demonstrates effectiveness across single-task, large-scale, and multi-task scenarios.
• The proposed MoE architecture shows promise for advancing GLAD research.

Conclusions:

• MoEGAD offers a robust solution for GLAD, particularly in low-data regimes.
• The framework's adaptability makes it suitable for diverse and complex GLAD problems.
• This work pioneers the application of MoE architectures in graph-level anomaly detection.