Graph-level anomaly detection (GLAD) aims to identify graphs that deviate markedly from the majority. Existing solutions usually follow a dataset‑specific paradigm, requiring abundant in‑domain labels and heavy training, which limits their deployment in real scenarios where labeled data are scarce and new domains constantly appear. This paper studies generalist GLAD, i.e., learning a single model that can detect anomalous graphs across multiple domains with minimal supervision on target data. The problem brings two key challenges: extracting transferable anomaly patterns from heterogeneous graph collections and adapting efficiently under limited supervision. To tackle them we propose GenGLAD. The approach centers on a graph neural network‑based discriminative subgraph extraction model that learns substructures distinguishing anomalous from normal graphs across diverse domains. At test time, multi‑dimensional discrepancies between the extracted subgraph and multi‑level contextual information are computed to characterize anomalies from several perspectives. A lightweight scoring module then automatically selects the most informative signals, improving robustness against misleading patterns. Extensive experiments on various benchmarks demonstrate that GenGLAD achieves strong performance, superior generalization, and higher data efficiency compared with prior methods.
Review: The paper presents a compelling unified framework for cross‑domain graph anomaly detection, showing that discriminative subgraph learning can effectively bridge domain gaps and operate with limited supervision.