Identifying Anomalous Interactions in Evolving Networks via Coupled Spatio-Temporal Encoding and Twin Diffusion Models
Lists
Only you see this
Abstract
Detecting malicious or abnormal edges within time-evolving networks remains a persistent challenge for security, financial, and social infrastructures. Most current representation learning approaches process network topology and temporal dynamics sequentially or independently, inadvertently discarding crucial synergistic patterns. Compounding this issue is the inherent sparsity of ground-truth anomalies and the widespread reliance on noisy surrogate labels, which severely degrade the reliability of decision boundaries. To overcome these limitations, we introduce a new detection framework that strictly couples structural and temporal features while employing a generative scoring paradigm. Specifically, we design a Joint Spatiotemporal Graph Encoder that leverages position-aware attention mechanisms to project intricate topological and time-step dependencies into a single, highly expressive embedding space. Building on these representations, we develop a Twin Diffusion Anomaly Scorer. Rather than forcing a hard discriminative boundary, this module models the generative trajectories of normative and anomalous edge distributions as two calibrated diffusion anomaly scoring processes. We further integrate a noise-rectification objective to mitigate the impact of unreliable pseudo-labels, ultimately deriving robust anomaly scores through a Bayesian probabilistic formulation. Empirical evaluations across diverse real-world temporal networks indicate that our framework yields substantial and consistent performance
BibTeX
@misc{yu2026identifying,
author = {Hang Yu},
title = {{Identifying Anomalous Interactions in Evolving Networks via Coupled Spatio-Temporal Encoding and Twin Diffusion Models}},
year = {2026},
url = {https://openreview.net/forum?id=6TPSRKi32s},
howpublished = {OpenReview},
}Loading related content...