Anomaly detection on graphs focuses on identifying irregular patterns or anomalous nodes within graph-structured data, which deviate significantly from the norm. This domain gains paramount importance due to its wide applicability in various fields such as spam detection, anti-money laundering, and network security. In the application of anomaly detection on graphs, tackling the challenges posed by label imbalance and data insufficiency is of significance. Recent proliferation in generative models, especially diffusion models, paves a promising way. In this paper, we introduce a graph diffusion model in latent space, designed to alleviate the label imbalance problem prevalent in anomaly detection on graphs. The proposed model is capable of multitask generation of graph structures and node features, and further endowed with conditional generative capabilities to produce only positive examples, thereby mitigating label imbalance issues. We improved the diffusion model to apply on both homogeneous graphs and heterogeneous graphs. Through extensive experiments, we demonstrate that our proposed method offers notable improvements over conventional techniques.
Research areas
