Tri-Level Anomaly Detection in Text-Attributed Graphs with Graph Foundation Model

Tri-Level Anomaly Detection in Text-Attributed Graphs with Graph Foundation Model

Moiseykin A.D. (ISP RAS, Moscow, Russia; MSU, Moscow, Russia)
Drobyshevskiy M.D. (ISP RAS, Moscow, Russia; MIPT, Dolgoprudny, Moscow Region, Russia)
Soloviov M.V. (ISP RAS, Moscow, Russia; MIPT, Dolgoprudny, Moscow Region, Russia)

Abstract

Text-attributed graphs (TAGs) have become increasingly important in modern applications such as citation networks, knowledge graphs, and social platforms. As their scale and complexity continue to grow, anomaly detection in TAGs has attracted substantial attention. Despite recent progress, effectively exploiting the heterogeneous information available in these graphs remains a challenging task. At the same time, graph foundation models have recently emerged as a promising direction in graph learning, yet their potential for anomaly detection is still underexplored. In this paper, we propose the Tri-Level Anomaly Detection model (TriAD), a novel model for text-attributed graph anomaly detection that jointly leverages three levels of information: graph structure, node text, and representations obtained from a pretrained graph foundation model. TriAD aligns these levels in a shared representation space via cross-level contrastive learning, allowing it to capture inconsistencies between local graph context, semantic content, and higher-order structural patterns. By incorporating foundation-model-based graph representations, TriAD provides an additional anomaly signal that complements text and local graph structure. Across four benchmarks, TriAD obtains the best ROC-AUC on all four and the best Average Precision on three. In addition, we introduce an anomaly generation procedure to construct a new WordNet-domain dataset with a broader variety of synthetic anomalies. This dataset may be useful for future studies on anomaly detection in text-attributed graphs.

Keywords

graph anomaly detection; graph foundation models; text-attributed graphs; contrastive learning.

Edition

Proceedings of the Institute for System Programming, vol. 38, issue 6, part 1, 2026, pp. 73-86

ISSN 2220-6426 (Online), ISSN 2079-8156 (Print).

DOI: 10.15514/ISPRAS-2026-38(6)-5

For citation

Moiseykin A.D., Drobyshevskiy M.D., Soloviov M.V. Tri-Level Anomaly Detection in Text-Attributed Graphs with Graph Foundation Model. Proceedings of the Institute for System Programming, vol. 38, issue 6, part 1, 2026, pp. 73-86 DOI: 10.15514/ISPRAS-2026-38(6)-5.

Full text of the paper in pdf Back to the contents of the volume