Calibrated Diffusion and Spatio-Temporal Unified Representation for Robust Graph Anomaly Detection

Chaoqi Chen

Abstract: The proliferation of dynamic graph-based applications in financial transaction systems, social networks, and e-commerce platforms has made anomaly edge detection an increasingly critical security challenge. However, current approaches suffer from two fundamental limitations: they inadequately capture joint spatio-temporal dependencies in edge representations, and they fail to robustly model anomalous distributions when learning from noisy pseudo-labels. To address these challenges, we present a comprehensive framework that integrates advanced representation learning with robust generative modeling. Our module employs unified spatio-temporal encoding on spatio-temporal graphs and introduces position-aware attention mechanisms to construct highly expressive edge representations that preserve both structural context and temporal evolution. To overcome the challenge of noisy supervision inherent in pseudo-anomalous edge generation, we propose CDAS module, which implements a calibrated dual diffusion framework in latent space. By optimizing a novel calibration loss that explicitly rectifies noise in pseudo-anomalous labels, our approach ensures discriminative modeling of anomalous distributions. Extensive experiments across five real-world datasets—including Bitcoin-alpha, Bitcoin-OTC, UCI Messages, Digg, and Email-DNC—demonstrate that Our method consistently achieves superior performance compared to state-of-the-art baselines, with AUC improvements exceeding 7% on financial networks and 3% on social networks.