Noise-Calibrated Dual Latent Diffusion for Anomalous Edge Detection with Unified Spatio-Temporal Graph Encoding

Luyao Tang, Kunze Huang, Jiewei Zheng

Abstract: Anomalous edge detection in dynamic graphs underpins high-impact applications such as fraud monitoring in financial networks, abuse detection in social platforms, and risk control in e-commerce. Two obstacles continue to limit practical accuracy: (i) edge representations often miss joint spatio-temporal correlations because topology and time are encoded in a decoupled manner, and (ii) anomaly modeling is unreliable when pseudo-anomalous edges produced by negative sampling contain substantial label noise. We propose ST-DiffDyG, a dynamic graph anomaly detection framework that jointly resolves these issues through representation learning and calibrated generative scoring. On the representation side, the module constructs spatio-temporal graphs and applies unified spatio-temporal encoding together with position-aware spatio-temporal attention to produce expressive edge embeddings. On the detection side, the secend module introduces a calibrated dual diffusion design in latent space, where separate diffusion models characterize normal and pseudo-anomalous edge distributions and a calibration loss leverages an additional holistic diffusion model to correct the bias induced by noisy pseudo labels. Anomaly scores are then computed via Bayesian inference over the learned distributions. Experiments on multiple real-world benchmarks demonstrate that ST-DiffDyG consistently surpasses state-of-the-art baselines under severe class imbalance.