AELAD: Adaptive Evidential Learning for Anomaly Detection in Dynamic Graphs

Dynamic graphs are pervasive in real-world scenarios such as financial transactions and social networks, making the detection of anomalous nodes a problem of significant practical importance. However, existing methods predominantly adopt a point estimation paradigm. They lack uncertainty awareness and cannot quantify the reliability of anomaly predictions under sparse histories or highly variable node behaviors. Furthermore, feature-based anomalies and structure-based anomalies are inherently heterogeneous signals whose naive fusion degrades detection performance due to inconsistent scales. To address these challenges, we propose AELAD, a framework for anomaly detection in dynamic graphs based on adaptive evidential learning. Specifically, AELAD employs an evidential learning mechanism that enables the model to output distributional parameters rather than point estimates. Moreover, a Normal-Inverse Gamma (NIG) distribution and a Beta distribution independently model feature and topological uncertainty, and the resulting scores are adaptively fused via learnable weights. Extensive experiments on six benchmark datasets demonstrate that AELAD achieves superior anomaly detection performance compared to state-of-the-art baselines.