AnomalyGFM: Graph Foundation Model for Zero/Few-shot Anomaly Detection

Hezhe Qiao; Chaoxi Niu; Ling Chen; Guansong Pang

cs.LGFeb 13, 2025

AnomalyGFM: Graph Foundation Model for Zero/Few-shot Anomaly Detection

Authors: Hezhe Qiao, Chaoxi Niu, Ling Chen, Guansong Pang

Organizations: School of Computing and Information Systems Singapore Management University Singapore · Australian Artificial Intelligence Institute University of Technology Sydney Sydney, Australia

Abstract

Graph anomaly detection (GAD) aims to identify abnormal nodes that differ from the majority of the nodes in a graph, which has been attracting significant attention in recent years. Existing generalist graph models have achieved remarkable success in different graph tasks but struggle to generalize to the GAD task. This limitation arises from their difficulty in learning generalized knowledge for capturing the inherently infrequent, irregular and heterogeneous abnormality patterns in graphs from different domains. To address this challenge, we propose AnomalyGFM, a GAD-oriented graph foundation model that supports zero-shot inference and few-shot prompt tuning for GAD in diverse graph datasets. One key insight is that graph-agnostic representations for normal and abnormal classes are required to support effective zero/few-shot GAD across different graphs. Motivated by this, AnomalyGFM is pre-trained to align data-independent, learnable normal and abnormal class prototypes with node representation residuals (i.e., representation deviation of a node from its neighbors). The residual features essentially project the node information into a unified feature space where we can effectively measure the abnormality of nodes from different graphs in a consistent way. This provides a driving force for the learning of graph-agnostic, discriminative prototypes for the normal and abnormal classes, which can be used to enable zero-shot GAD on new graphs, including very large-scale graphs. If there are few-shot labeled normal nodes available in the new graphs, AnomalyGFM can further support prompt tuning to leverage these nodes for better adaptation. Comprehensive experiments on 11 widely-used GAD datasets with real anomalies, demonstrate that AnomalyGFM significantly outperforms state-of-the-art competing methods under both zero- and few-shot GAD settings.

Figures & tables

Figure 1. Given a set of graph datasets consisting of normal nodes (blue) and abnormal nodes (red) as in (a) , AnomalyGFM aims to learn graph-agnostic prototypes for normal and abnormal nodes in a projected residual feature space. Compared to the original feature space where normal and abnormal patterns are irregularly distributed and differ significantly in different datasets as in (b) , the residual feature space learned by AnomalyGFM in (c) is universally effective in separating anomalies from normal nodes in different graphs. In (b) and (c), AnomalyGFM is trained on Facebook ( Xu et al., 2022 ) and evaluated directly on Amazon ( Dou et al., 2020 ) and Question ( Platonov et al., 2023 ) ; both feature spaces are learned by AnomalyGFM.

Figure 2. Overview of AnomalyGFM. (a) During the pre-training, AnomalyGFM aims to align two learnable class-level prototypes with the representation residuals of normal and abnormal nodes to learn discriminative graph-agnostic prototypes on an auxiliary labeled graph. (b) During inference, the anomaly score is determined in a consistent way using similarity between node representation residual and prototypes where all learnable parameters are frozen. Given a target graph with few-shot labeled normal nodes, these nodes are used to better adapt the pre-trained normal class prototypes to the target graph through a one-class prompt tuning method.

Dataset	Domain	#Nodes	#Edges	#Feat.	Ano.	Sim.
Facebook	Social Networks	1,081	55,104	576	25(2.31%)	0.690
Reddit	Social Networks	10,984	168,016	64	366(3.33%)	0.997
Amazon	Co-review	10,244	175,608	25	693(6.66%)	0.645
Disney	Co-purchase	124	335	28	6(4.8%)	0.804
Amazon-all	Co-review	11,944	4,398,392	25	821(6.87%)	0.645
YelpChi-all	Co-review	45,941	3,846,979	32	6,674(14.52%)	0.905

Table 1. Key statistics of GAD datasets used in our experiments. Ano. indicates the number (rate) of anomalies in the graph and Sim. indicates the global average edge similarity.

Metric	Method	Dataset									Avg.	p-value
Metric	Method	Reddit	Amazon	Disney	Amazon-all	YelpChi-all	Tolokers	Question	Elliptic	T-Finance	Avg.	p-value
AUROC	Unsupervised Methods
	AnomalyDAE (ICASSP’20)	0.5016	0.5818	0.4853	0.7228	0.5002	0.5948	0.4311	0.4197	0.2324	0.4966	0.007
	CoLA (TNNLS’21)	0.4623	0.4580	0.4696	0.4091	0.4879	0.4501	0.4945	0.5572	0.4889	0.4752	0.003
	TAM (NeurIPS’23)	0.5725	0.4720	0.4773	0.7543	0.4216	0.5351	0.5119	0.3282	0.2990	0.4857	0.003
	GADAM (ICLR’24)	0.4532	0.6646	0.4288	0.5959	0.4829	0.4832	0.5594	0.3922	0.1382	0.4664	0.007

Table 2. AUROC and AUPRC results for zero-shot GAD on nine real-world datasets, with the models trained on Facebook only. For each dataset, the best performance per column within each metric is boldfaced, with the second-best underlined. Avg. denotes the average performance. P-values are from one-tailed Wilcoxon signed rank tests.

Metric	Setting	Method	Dataset									Avg.	p-value
Metric	Setting	Method	Reddit	Amazon	Disney	Amazon-all	YelpChi-all	Tolokers	Question	Elliptic	T-Finance	Avg.	p-value
AUROC	1-shot	GPPT (KDD’22)	0.5000	0.5303	0.4997	0.5010	0.5000	0.5061	0.4921	0.6162	0.3647	0.5011	0.003
		GraphPrompt (WebConf’23)	0.4216	0.4882	0.4223	0.2631	0.4811	0.5328	0.4086	0.6001	0.4000	0.4464	0.004
		ARC (NeurIPS’24)	0.4899	0.4571	0.3578	0.4570	0.4910	0.4667	0.5865	0.2904	0.2484	0.4272	0.008
		AnomalyGFM	0.5922	0.8531	0.6649	0.8972	0.5872	0.5898	0.5303	0.6199	0.5916	0.6584	/
	5-shot	GPPT (KDD’22)	0.5000	0.5098	0.5000	0.5051	0.5000	0.5181	0.4959	0.5736	0.2609	0.4848	0.003

Table 3. Results on the nine real-world GAD datasets under 1/5/10-shot with models tuned on the few-shot data.

AUROC	Unsupervised Methods
Metric	Method	Dataset
Metric	Method	T-Finance	T-Social
	TAM (NeurIPS’23)	0.2990	/
	GADAM (ICLR’24)	0.1382	0.5155
	Supervised Methods
	BWGNN (ICML’22)	0.5457	0.4964

Table 4. Subgraph inference on the very large-scale graphs. ‘/’ indicates that the model cannot handle the dataset.

Metric	Method	Dataset
Metric	Method	Reddit	Amazon	Disney	Amazon-all	YelpChi-all	Tolokers	Question	Elliptic	T-Finance	Avg.
AUROC	FA	0.5697	0.5680	0.6073	0.5097	0.5211	0.5594	0.5326	0.7765	0.6674	0.5901
	BCE	0.5445	0.4308	0.7542	0.5545	0.5232	0.5187	0.4593	0.2581	0.6581	0.5223
	BCE-R	0.5108	0.5314	0.4887	0.5777	0.4590	0.4788	0.4601	0.5414	0.5380	0.5095
	AnomalyGFM	0.5974	0.8417	0.6751	0.9032	0.5791	0.5843	0.5380	0.6195	0.5614	0.6555
AUPRC	FA	0.0400	0.0776	0.0960	0.0684	0.1555	0.2478	0.0356	0.1199	0.1495	0.1100

Table 5. Ablation study results for AnomalyGFM and its variants.

Figure 3. AnomalyGFM performance w.r.t subgraph size $s$ .

Figure 4. AnomalyGFM performance w.r.t prototype size $T$ .

Figure 5. AnomalyGFM performance w.r.t $\alpha$ .

Figure 6. AUROC and AUPRC of AnomalyGFM w.r.t $d^{\prime}$

Figure 7. AUROC and AUPRC of AnomalyGFM w.r.t $\beta$ .

Methods	AnomalyDAE	TAM	GAT	BWGNN	UNPrompt	AnomalyGFM
Training Time	86.04	479.70	2.43	4.86	2.08	7.45
Inference Time	170.46	252.52	91.88	98.33	19.47	11.00

Table 6. Training and inference time (seconds) comparison.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Metric	Setting	Method	Dataset
Metric	Setting	Method	Amazon-all	YelpChi-all
AUROC	1-shot	BWGNN (ICML’22)	0.7441	0.5447
		MetaGAD (DSAA’24)	0.8602	0.5434
		AnomalyGFM	0.9155	0.5941
	5-shot	BWGNN (ICML’22)	0.8292	0.4759
		MetaGAD (DSAA’24)	0.8558	0.5741

Appendix

Table 7. Comparison under 1/5-shot anomalies

AUROC	Unsupervised Methods
Metric	Method	Dataset								Avg.	p-value
Metric	Method	Facebook	Reddit	Disney	YelpChi-all	Tolokers	Question	Elliptic	T-Finance	Avg.	p-value
	AnomalyDAE (ICASSP’20)	0.6123	0.5799	0.4938	0.4912	0.5080	0.5401	0.2769	0.2886	0.4738	0.0156
	CoLA (TNNLS’21)	0.5427	0.4962	0.5455	0.4937	0.4680	0.4768	0.6654	0.4108	0.5123	0.0390
	GADAM (ICLR’24)	0.6024	0.4720	0.3966	0.5289	0.5193	0.5575	0.2907	0.1463	0.4391	0.0234
	Supervised Methods

Appendix

Table 8. Comparison of AUROC and AUPRC results on 8 real-world GAD datasets under zero-shot setting with the models trained on Amazon only.

Setting	Sim. >0.5	Sim. <=0.5
Zero-shot Setting	0	4
Few-shot Setting	0.5	4

Appendix

Table 9. The value of

\beta

in anomaly scoring in both zero-shot and few-shot setting. Sim. indicate the global average edge similarity of the graph.

Metric	Initialization	Dataset
Metric	Initialization	Amazon-all	Tolokers	Question	Elliptic	T-Finance
AUROC	$\mu$ = 0, $\sigma$ = 1	0.9032	0.5843	0.5280	0.6195	0.5614
	$\mu$ = 0, $\sigma$ = 0.5	0.9229	0.5981	0.5378	0.6068	0.5902
	$\mu$ = 0.1, $\sigma$ = 0.5	0.8756	0.6004	0.5437	0.5907	0.5727
	$\mu$ = 0.2, $\sigma$ = 1	0.8963	0.5975	0.5429	0.6099	0.5922
AUPRC	$\mu$ = 0, $\sigma$ = 1	0.6820	0.2749	0.0335	0.1333	0.0593

Appendix

Table 10. Performance of AnomalyGFM under different initialization hyperparameters for Gaussian distribution

Graph-Based Anomaly Detection

See all 37 papers

Feb 21, 2026

From Few-Shot to Zero-Shot: Towards Generalist Graph Anomaly Detection

Graph anomaly detection (GAD) is critical for identifying abnormal nodes in graph-structured data from diverse domains, including cybersecurity and social networks. The existing GAD methods often focus on the learning paradigms of "one-model-for-one-dataset", requiring dataset-specific training for each dataset to achieve optimal performance. However, this paradigm suffers from significant limitations, such as high computational and data costs, limited generalization and transferability to new datasets, and challenges in privacy-sensitive scenarios where access to full datasets or sufficient labels is restricted. To address these limitations, we propose a novel generalist GAD paradigm that aims to develop a unified model capable of detecting anomalies on multiple unseen datasets without extensive retraining/fine-tuning or dataset-specific customization. To this end, we propose ARC, a few-shot generalist GAD method that leverages in-context learning and requires only a few labeled normal samples at inference time. Specifically, ARC consists of three core modules: a feature Alignment module to unify and align features across datasets, a Residual GNN encoder to capture dataset-agnostic anomaly representations, and a cross-attentive in-Context learning module to score anomalies using few-shot normal context. Building on ARC, we further introduce ARC_zero for the zero-shot generalist GAD setting, which selects representative pseudo-normal nodes via a pseudo-context mechanism and thus enables fully label-free inference on unseen datasets. Extensive experiments on 17 real-world graph datasets demonstrate that both ARC and ARC_zero effectively detect anomalies, exhibit strong generalization ability, and perform efficiently under few-shot and zero-shot settings.

Yixin Liu, Shiyuan Li, Yu Zheng +4

School of Information and Communication Technology, Griffith University, Australia. · School of Computer, Electronics and Information, Guangxi University, Nanning, China. · Department of Data Science and Artificial Intelligence, The Hong Kong Polytechnic University, Hong Kong. +1

Anomaly Detection Zero-Shot Anomaly Detection Few-Shot Anomaly Detection

Ask PDF

Sep 16, 2026

FoundAna: A GNN-assisted Foundation Model for Graph Anomaly Detection

Graph anomaly detection aims to identify graph structures (e.g., nodes, edges, or subgraphs) that deviate significantly from expected patterns, which supports critical applications in fraud detection, spam identification, network intrusion, etc. Despite the growing methods in the field, existing approaches follow a one-model-per-dataset paradigm, limiting their transferability across diverse real-world scenarios due to task heterogeneity, label scarcity, and domain variability. In this work, we introduce FoundAna, a GNN-assisted Foundation Model for Graph Anomaly Detection - the first foundation model framework designated for generalizable, cross-graph anomaly detection by combining GNNs and transformers. FoundAna integrates an anomaly detection-specific GNN component with a standard transformer encoder augmented by four complementary positional encodings, which enable the model to capture both local and global structural information. Specifically, the positional encoding enriched node representations are passed through attribute and adjacency decoders, and the reconstruction errors serve as the anomaly score. Extensive experiments on nine benchmark datasets spanning financial, social, and citation network domains demonstrate that FoundAna consistently outperforms state-of-the-art baselines. The code implementation and Supplementary materials are here: https://github.com/FoundAna331/FoundAna.

Suprim Nakarmi, Chahana Dahal, Yue Zhao +2

Department of Computer Science, University of Nevada Las Vegas, Las Vegas, USA · Department of Computer Science, University of Southern California, Los Angeles, USA

Graph Neural Networks Reconstruction-Based Anomaly Detection Graph Foundation Models Anomaly Detection Graph-Based Anomaly Detection

Ask PDF

May 26, 2026

Generalist Graph Anomaly Detection via Prototype-Based Distillation

Driven by the pressing demand for graph anomaly detection (GAD) in high-stakes domains, the generalist GAD paradigm, which trains a single detector transferable across new graphs, has recently gained growing attention. However, existing methods often rely on scarce and costly annotations for training and sometimes even require few-shot support at inference, which limits their robustness to diverse and unseen anomaly patterns. To address this limitation, we introduce ProMoS, the first unsupervised generalist GAD framework, which detects anomalies by modeling the abundant normality in unlabeled data. ProMoS adopts a knowledge-distillation paradigm to distill normality priors from a frozen self-supervised graph neural network (GNN) teacher to a mixture-of-students model with shared global and lightweight personalized branches, enabling efficient and expressive normality modeling without learning from scratch. We further propose prototype-guided soft-label distillation to align teacher and student in a shared prototype space, enhancing cross-graph generalizability. During inference, ProMoS performs zero-shot anomaly detection on unseen graphs via distillation bias and prototype geometric deviation. Extensive experiments show the effectiveness and efficiency of ProMoS, charting a practical path toward label-free, zero-shot generalist GAD.

Yiming Xu, Zihan Chen, Zhen Peng +4

School of Computer Science and Technology, Xi’an Jiao-tong University, Xi’an, China · National Engineering Research Center for Visual Information and Applications, Xi’an, China · University of Virginia, Charlottesville, USA +3

Unsupervised Anomaly Detection Zero-Shot Anomaly Detection Knowledge Distillation Anomaly Detection Graph-Based Anomaly Detection

Ask PDF

Abstract

Graph-Based Anomaly Detection

Explore similar work

From Few-Shot to Zero-Shot: Towards Generalist Graph Anomaly Detection

FoundAna: A GNN-assisted Foundation Model for Graph Anomaly Detection

Generalist Graph Anomaly Detection via Prototype-Based Distillation