Anomaly Detection through Conditional Diffusion Probability Modeling on Graphs
Keywords: Anomaly detection, Graph Neural Network, Diffusion Model
Abstract: Existing Graph Neural Network-based anomaly detection methods suffer from over-smoothing issues during feature aggregation. Moreover, most existing methods are discriminative models that learn the boundaries between anomalous and normal data points, allowing malicious nodes in a dynamic adversarial environment to bypass detection boundaries. To address these issues, existing methods primarily focus on enhancing the discriminative boundary for each individual node, rather than considering the interdependencies of node anomalies from a holistic graph perspective. We propose an advanced Conditional Graph Anomaly Diffusion Model (CGADM) to model and capture the joint distribution of anomalies on the whole graph, thereby enabling generative graph anomaly detection. To avoid starting the diffusion process from a random state, CGADM introduces a prior-guided denoising diffusion probability model. To circumvent the need for iterative denoising samplings for each node on large-scale graphs, we adopt a prior confidence-aware mechanism to dynamically adjust the reverse sampling steps for each node, significantly reducing the computational burden on large-scale graphs. We conducted experiments on CGADM using standard benchmarks, and the results demonstrated excellent performance in graph anomaly detection tasks. Additional ablation studies confirmed our framework's computational advantages.
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 6482
Loading