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

A fundamental challenge in graph-level anomaly detection (GLAD) is the scarcity of anomalous graph data, as the training dataset typically contains only normal graphs or very few anomalies. This imbalance hinders the development of robust detection models. In this paper, we propose Anomalous Graph Diffusion (AGDiff), a framework that explores the potential of diffusion models in generating pseudo-anomalous graphs for GLAD. Unlike existing diffusion-based methods that focus on modeling data normality, AGDiff leverages the latent diffusion framework to incorporate subtle perturbations into graph representations, thereby generating pseudo-anomalous graphs that closely resemble normal ones. By jointly training a classifier to distinguish these generated graph anomalies from normal graphs, AGDiff learns more discriminative decision boundaries. The shift from solely modeling normality to explicitly generating and learning from pseudo graph anomalies enables AGDiff to effectively identify complex anomalous patterns that other approaches might overlook. Comprehensive experimental results demonstrate that the proposed AGDiff significantly outperforms several state-of-the-art GLAD baselines.

Graph anomaly detection aims to identify graph-structured individuals that deviate from common patterns observed in normal cases. However, this task is particularly challenging due to the scarcity or even complete absence of anomalous examples in the training set. For instance, in early-stage cancer, a few abnormal cells may be hidden among many healthy ones and show no symptoms, yet early detection is vital for effective treatment. Hence, this paper introduces a method that leverages diffusion models to generate synthetic graph anomalies. Rather than learning only from normal data, it generates slightly perturbed graph structures that resemble anomalies and trains a model to distinguish them from normal graphs. This approach enhances the model’s ability to detect subtle and complex anomalies.