New recipes for graph anomaly detection: Forward diffusion dynamics and graph generation
Primary Area: learning on graphs and other geometries & topologies
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Graph anomaly detection, anomaly detection, denoising diffusion, graph generation
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Distinguishing atypical nodes in a graph, which is known as graph anomaly detection, is more crucial than the generic node classification in real applications, such as fraud and spam detection. However, the lack of prior knowledge about anomalies and the extremely class-imbalanced data pose formidable challenges in learning the distributions of normal nodes and anomalies, which serves as the foundation of the state of the arts. We introduce a novel paradigm (first recipe) for detecting graph anomalies, stemming from our empirical and rigorous analysis of the significantly distinct evolving patterns between anomalies and normal nodes when scheduled noise is injected into the node attributes, referred to as the forward diffusion process. Rather than modeling the data distribution, we present three non-GNN methods to capture the evolving patterns and achieve promising results on six widely-used datasets, while mitigating the oversmoothing limitation and shallow architecture of GNN methods. We further investigate the generative power of denoising diffusion models to synthesize training samples that align with the original graph semantics (second recipe). In particular, we derive two principles for designing the denoising neural network and generating graphs. With our proposed graph generation method, we attain record-breaking performance while our generated graphs are also capable of enhancing the results of existing methods. All the code and data are available at \url{https://github.com/DiffAD/DiffAD}.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7059
Loading