Reproduce “GTHNA: Local-global Graph Transformer with Memory Reconstruction for Holistic Node Anomaly Evaluation” on OpenResearch

About this paper

Mingkang Li, Xuexiong Luo, Yue Zhang, Yaoyang Li, Fu Lin

GTHNA combines local neighborhood extraction, global graph attention, and a selectively updated memory of normal-node prototypes for unsupervised node anomaly detection. Across seven benchmarks, it achieves the best reported AUC on six datasets, reaching 93.62% on Citeseer, 89.81% on ACM, 86.88% on YelpRes, and 90.41% on Amazon, while remaining below DiffGAD on Weibo.

Key insights

• Selective memory updates provide an explicit normality model and reduce—but do not eliminate—the risk that anomalous nodes contaminate reconstruction training.
• Neighborhood-level statistics complement attribute and edge reconstruction, enabling the detector to identify anomalies expressed through broader local distributions rather than individual features alone.
• Ablations show that both the local structure extractor and memory module improve performance, with the memory component generally producing the larger gain, especially on injected-anomaly datasets.

Summary by alphaXiv.