Title: Developing a Discriminative and Generalizable Anomaly Detector for Dynamic Graphs Under Limited Supervision

Abstract:

Although dynamic graph anomaly detection is essential for numerous practical applications, it faces significant hurdles due to the lack of labeled anomalous instances. Current approaches typically fall into unsupervised or semi-supervised categories. While unsupervised techniques eliminate the dependency on labeled data, they frequently struggle with unclear decision boundaries. Conversely, semi-supervised models tend to overfit the small set of available labeled anomalies, resulting in poor generalization to novel, unseen anomalies. To bridge this gap, we investigate a largely overlooked challenge: establishing a discriminative boundary using normal and unlabeled data, while effectively utilizing limited labeled anomalies when they are present, without compromising the model's ability to generalize to new anomalies.

In this work, we introduce a robust, model-agnostic framework designed to be both effective and generalizable. The framework comprises three key elements: (i) a residual representation encoding mechanism that identifies deviations between current interactions and their historical context, thereby generating signals relevant to anomalies; (ii) a restriction loss function that confines normal representations within an interval defined by two co-centered hyperspheres, which maintains consistent scales while ensuring anomalies remain distinct; and (iii) a bi-boundary optimization strategy that employs a normalizing flow to model the log-likelihood distribution, facilitating the learning of a discriminative and resilient boundary. Our extensive experimental results confirm the superior performance of our proposed framework across a variety of evaluation scenarios.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC