Modeling Epistemic Uncertainty for Anomalous Link Discovery
Lists
Only you see this
Abstract
Streaming relational data—from denial-of-log alerts to friendship requests—poses a fundamental challenge: how do we flag suspicious edges when the underlying distribution itself drifts, and ground-truth labels are noisy pseudo-labels and class imbalance? We present BayesDyG, a detection pipeline that reframes dynamic graph edges as probabilistic paths through spatio-temporal unified graph representation. Rather than decoupling spatial and temporal features which losing joint correlations, we maintain two concurrent score functions derived from partially noised diffusion trajectories. Structural-Temporal Fusion replaces rigid snapshot separation with continuous-time message passing augmented by Hawkes-process intensity decay; this preserves transductive signals that vanish in windowed batching. The second component, calibrated diffusion anomaly scoring addresses the reality that human annotators frequently mistake outliers for anomalies. By interpreting pseudo-labels as noisy observations of a hidden Bernoulli process, we derive an evidence-lower-bound that down weights uncertain points during the reverse diffusion phase. Empirically, BayesDyG recovers more ground-truth fraud cases across six benchmarks than with dynamic graph anomaly detection methods that actually correlate with accuracy.
BibTeX
@misc{ding2026modeling,
author = {Xinghao Ding},
title = {{Modeling Epistemic Uncertainty for Anomalous Link Discovery}},
year = {2026},
url = {https://openreview.net/forum?id=m2UZlrpGLW},
howpublished = {OpenReview},
}Loading related content...