Dual Diffusion for Dynamic Graph Anomaly Detection Abstract

Luyao Tang, Zheyuan Cai, Jiewei Zheng

Abstract:

Edge-level anomaly detection on evolving graphs must distinguish rare irregular interactions from legitimate but rapidly changing behavior. Existing approaches often fall into one of two traps: they either learn overly permissive decision boundaries when only normal edges are observed, or they depend on pseudo-supervision from negative sampling and become sensitive to mislabeled pseudo anomalies. At the same time, common temporal encoders may blur spatio-temporal signals when topology and time are processed in loosely coupled stages.
We introduce a method that couples a spatio-temporal unified representation module with a calibrated dual diffusion anomaly scoring module. The representation component constructs spatio-temporal context sequences and employs unified spatio-temporal encoding plus position-aware attention to generate expressive edge representations that preserve structural and temporal cues. The scoring component performs latent-space diffusion modeling with two diffusion models (normal vs. pseudo-anomalous) and optimizes a novel calibration loss guided by a holistic diffusion model to correct the inherent noise in pseudo-anomalous labels. Using Bayesian inference to compute anomaly scores, the method yields more discriminative detection. Evaluations across real-world dynamic graph benchmarks show improved AUC and Macro-F1 compared with state-of-the-art alternatives.