Dual-Process Diffusion with Unified Temporal-Structural Encoding

Yue Yao

Abstract: Anomaly detection in dynamic networks requires modeling both evolving topology and temporal interaction patterns, yet prior methods often fail to integrate these aspects effectively. Moreover, the reliance on imperfect pseudo-anomaly labels and skewed class distributions makes it difficult to obtain reliable decision functions. We present a novel approach that addresses these challenges through a unified temporal-structural encoding scheme coupled with a calibrated diffusion framework. Our model constructs spatio-temporal unified graph representation by jointly embedding temporal sequences and graph structure, guided by an attention mechanism that is sensitive to positional and temporal context. On top of this representation, we develop a two-branch diffusion process that separately characterizes typical and atypical behaviors, allowing clearer discrimination between them. To reduce the impact of noisy supervision, we introduce a calibration strategy that adjusts the learning signals during training. The resulting calibrated diffusion anomaly scoring are computed using a Bayesian perspective that fuses information from both diffusion branches. Empirical evaluations across multiple datasets show that the proposed method achieves superior detection performance and improved stability compared to existing techniques.