Leveraging Diffusion Model as Pseudo-Anomalous Graph Generator for Graph-Level Anomaly Detection

Motivation 1. Unsupervised GLAD methods generally focus on modelling normal graph distributions, which struggles to identify subtle anomalies, especially those near the boundaries of normal graphs. 2. Semi-supervised GLAD methods can leverage limited labelled anomalies to enhance decision boundary learning. However, their effectiveness is constrained by the scarcity and diversity of labelled anomalous graph. 1. We introduce AGDiff, the first framework that explores the potential of diffusion models to mitigate the anomaly scarcity challenge in GLAD. 2. We propose a latent diffusion process with perturbation conditions to generate pseudo-anomalous graphs without relying on any labelled anomalies for improving decision boundary learning. 3. We demonstrate the effectiveness of AGDiff across extensive comparisons with state-of-the-art GLAD baselines on diverse graph benchmarks.