Learning Joint Spatio-Temporal Edge Representations for Graph Anomaly Detection
Lists
Only you see this
Abstract
Detecting anomalous edges in evolving graphs plays an essential role in domains such as online platforms, transaction networks, and cybersecurity. Many existing approaches treat structural and temporal signals separately, which limits their ability to capture coupled dependencies and often results in unstable anomaly boundaries under noisy supervision and severe class imbalance. In this work, we introduce a unified framework that jointly models spatial and temporal dynamics to improve edge-level anomaly detection. Specifically, we design a spatio-temporal unified graph representation module that encodes node interactions through a shared embedding space augmented with position-aware attention, enabling fine-grained modeling of temporal evolution and structural context. To further enhance robustness, we propose a dual diffusion-based scoring mechanism that learns normality and abnormality distributions in parallel. A calibrated diffusion anomaly scoring is incorporated to mitigate noise in pseudo labels and stabilize the learning process. Final anomaly scores CDAS are derived via a probabilistic inference scheme that integrates both diffusion processes. Experiments on several real-world dynamic graph benchmarks demonstrate consistent improvements over competitive baselines, highlighting the effectiveness of joint representation learning and calibrated diffusion modeling.
BibTeX
@misc{yao2026learning,
author = {Yue Yao},
title = {{Learning Joint Spatio-Temporal Edge Representations for Graph Anomaly Detection}},
year = {2026},
url = {https://openreview.net/forum?id=tVDrFOzk2r},
howpublished = {OpenReview},
}Loading related content...