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

Yixin Liu; Shiyuan Li; Yu Zheng; Qingfeng Chen; Chengqi Zhang; Philip S. Yu; Shirui Pan

cs.LGFeb 21, 2026

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

Authors: Yixin Liu, Shiyuan Li, Yu Zheng, Qingfeng Chen, Chengqi Zhang, Philip S. Yu, Shirui Pan

Organizations: 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. · Department of Computer Science, University of Illinois at Chicago, Chicago, IL 60607-7053, USA.

Abstract

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.

Figures & tables

Fig. 1: The sketch maps of the existing data-specific GAD paradigms ((a) supervised and (b) unsupervised) and our generalist GAD paradigms ((c) few-shot and (d) zero-shot).

Fig. 2: The overall pipeline of the proposed methods ARC, illustrating by (a)➜(b)➜(c), and ARC zero , illustrating by (a)➜(b)➜(d).

Panel of Figure 3 — Fig. 3: AUROC on data with 5 groups of feature subsets.

Fig. 4: Toy examples of reconstruction-based scoring mechanism. In each case, nodes 1, 2, 3, and 4 are normal samples, while node 5 is an anomaly.

Citation network with injected anomalies
Dataset	$\mathcal{T}_{train}$	$\mathcal{T}_{test}$	#Nodes	#Edges	#Feat.	#Ano.
Cora	-	$\checkmark$	2,708	5,429	1,433	150
CiteSeer	-	$\checkmark$	3,327	4,732	3,703	150
ACM	-	$\checkmark$	16,484	71,980	8,337	597
PubMed	$\checkmark$	-	19,717	44,338	500	600
Social network with injected anomalies

TABLE I: The statistics of datasets.

Method	Cora	CiteSeer	ACM	BlogCatalog	Facebook	Weibo	Reddit
Supervised - Pre-Train Only
GCN	$59.64{\scriptstyle\pm 8.30}$	$60.27{\scriptstyle\pm 8.11}$	$60.49{\scriptstyle\pm 9.65}$	$56.19{\scriptstyle\pm 6.39}$	$29.51{\scriptstyle\pm 4.86}$	$76.64{\scriptstyle\pm 17.69}$	$50.43{\scriptstyle\pm 4.41}$
GAT	$50.06{\scriptstyle\pm 2.65}$	$51.59{\scriptstyle\pm 3.49}$	$48.79{\scriptstyle\pm 2.73}$	$50.40{\scriptstyle\pm 2.80}$	$51.88{\scriptstyle\pm 2.16}$	$53.06{\scriptstyle\pm 7.48}$	$51.78{\scriptstyle\pm 4.04}$
BGNN	$42.45{\scriptstyle\pm 11.57}$	$42.32{\scriptstyle\pm 11.82}$	$44.00{\scriptstyle\pm 13.69}$	$47.67{\scriptstyle\pm 8.52}$	$54.74{\scriptstyle\pm 25.29}$	$32.75{\scriptstyle\pm 35.35}$	$50.27{\scriptstyle\pm 3.84}$
BWGNN	$54.06{\scriptstyle\pm 3.27}$	$52.61{\scriptstyle\pm 2.88}$	$67.59{\scriptstyle\pm 0.70}$	$56.34{\scriptstyle\pm 1.21}$	$45.84{\scriptstyle\pm 4.97}$	$53.38{\scriptstyle\pm 1.61}$	$48.97{\scriptstyle\pm 5.74}$
GHRN	$59.89{\scriptstyle\pm 6.57}$	$56.04{\scriptstyle\pm 9.19}$	$55.65{\scriptstyle\pm 6.37}$	$57.64{\scriptstyle\pm 3.48}$	$44.81{\scriptstyle\pm 8.06}$	$51.87{\scriptstyle\pm 14.18}$	$46.22{\scriptstyle\pm 2.33}$

TABLE II: Anomaly detection performance in terms of AUROC (in percent, mean

\pm

std). Highlighted are the results ranked first , second , and third . “Rank” indicates the average ranking over 13 datasets. Superscript “*” on a baseline method name indicates that our method is significantly different from this baseline on ALL_DATASETS_TRIAL_AVG (paired t-test and Wilcoxon test, p

<

0.05).

Variant	Cora	CiteSeer	ACM	BlogCatalog	Facebook	Weibo	Reddit	Amazon
ARC	${\color[rgb]{0.0781,0.5,0.332}\underline{\mathbf{87.45{\scriptstyle\pm 0.74}}}}$	${\color[rgb]{0.0781,0.5,0.332}\underline{\mathbf{90.95{\scriptstyle\pm 0.59}}}}$	${\color[rgb]{0.0781,0.5,0.332}\underline{\mathbf{79.88{\scriptstyle\pm 0.28}}}}$	${\color[rgb]{0.0781,0.5,0.332}\underline{\mathbf{74.76{\scriptstyle\pm 0.06}}}}$	${\color[rgb]{0.0781,0.5,0.332}\underline{\mathbf{67.56{\scriptstyle\pm 1.60}}}}$	$88.85{\scriptstyle\pm 0.14}$	${\color[rgb]{0.0781,0.5,0.332}\underline{\mathbf{60.04{\scriptstyle\pm 0.69}}}}$	${\color[rgb]{0.0781,0.5,0.332}\underline{\mathbf{80.67{\scriptstyle\pm 1.81}}}}$
w/o A	$80.65{\scriptstyle\pm 0.71}$	$83.35{\scriptstyle\pm 0.64}$	$79.29{\scriptstyle\pm 0.16}$	$73.86{\scriptstyle\pm 0.18}$	$62.80{\scriptstyle\pm 2.06}$	$89.69{\scriptstyle\pm 0.17}$	$54.60{\scriptstyle\pm 1.92}$	$64.76{\scriptstyle\pm 2.13}$
w/o R	$37.44{\scriptstyle\pm 1.40}$	$31.52{\scriptstyle\pm 0.71}$	$61.83{\scriptstyle\pm 1.16}$	$49.30{\scriptstyle\pm 2.06}$	$20.38{\scriptstyle\pm 9.63}$	${\color[rgb]{0.0781,0.5,0.332}\underline{\mathbf{97.72{\scriptstyle\pm 0.59}}}}$	$52.94{\scriptstyle\pm 0.96}$	$50.15{\scriptstyle\pm 0.24}$
w/o C	$47.39{\scriptstyle\pm 0.42}$	$53.98{\scriptstyle\pm 0.72}$	$54.24{\scriptstyle\pm 1.32}$	$60.46{\scriptstyle\pm 1.23}$	$48.86{\scriptstyle\pm 0.97}$	$42.84{\scriptstyle\pm 3.01}$	$51.03{\scriptstyle\pm 0.86}$	$69.02{\scriptstyle\pm 0.97}$

TABLE III: Performance of ARC and its variants in terms of AUROC.

Variant	Cora	CiteSeer	ACM	BlogCatalog	Facebook	Weibo	Reddit	Amazon
ARC zero	$87.54{\scriptstyle\pm 0.57}$	$91.20{\scriptstyle\pm 0.27}$	$78.78{\scriptstyle\pm 0.17}$	${\color[rgb]{0.0781,0.5,0.332}\underline{\mathbf{74.75{\scriptstyle\pm 0.03}}}}$	${\color[rgb]{0.0781,0.5,0.332}\underline{\mathbf{69.25{\scriptstyle\pm 0.95}}}}$	${\color[rgb]{0.0781,0.5,0.332}\underline{\mathbf{88.71{\scriptstyle\pm 0.18}}}}$	${\color[rgb]{0.0781,0.5,0.332}\underline{\mathbf{59.57{\scriptstyle\pm 0.42}}}}$	${\color[rgb]{0.0781,0.5,0.332}\underline{\mathbf{83.76{\scriptstyle\pm 1.01}}}}$
w/o Iter. Score	$87.76{\scriptstyle\pm 0.60}$	$90.13{\scriptstyle\pm 0.39}$	$79.23{\scriptstyle\pm 0.13}$	$74.71{\scriptstyle\pm 0.15}$	$67.78{\scriptstyle\pm 1.00}$	$88.17{\scriptstyle\pm 0.13}$	$59.13{\scriptstyle\pm 0.53}$	$80.85{\scriptstyle\pm 1.64}$
w/o Multi-Round	$87.37{\scriptstyle\pm 0.55}$	${\color[rgb]{0.0781,0.5,0.332}\underline{\mathbf{91.48{\scriptstyle\pm 0.26}}}}$	${\color[rgb]{0.0781,0.5,0.332}\underline{\mathbf{79.70{\scriptstyle\pm 0.09}}}}$	$74.47{\scriptstyle\pm 0.04}$	$65.52{\scriptstyle\pm 0.93}$	$88.25{\scriptstyle\pm 0.13}$	$58.16{\scriptstyle\pm 0.74}$	$80.92{\scriptstyle\pm 1.09}$
w Random Init	$86.47{\scriptstyle\pm 0.17}$	$90.10{\scriptstyle\pm 0.21}$	$79.69{\scriptstyle\pm 0.04}$	$74.15{\scriptstyle\pm 0.13}$	$65.40{\scriptstyle\pm 1.12}$	$79.36{\scriptstyle\pm 4.32}$	$58.92{\scriptstyle\pm 0.33}$	$82.91{\scriptstyle\pm 0.27}$
w Degree Init	$87.30{\scriptstyle\pm 0.52}$	$91.31{\scriptstyle\pm 0.23}$	$79.56{\scriptstyle\pm 0.09}$	$73.92{\scriptstyle\pm 0.17}$	$68.46{\scriptstyle\pm 1.62}$	$84.28{\scriptstyle\pm 2.18}$	$59.21{\scriptstyle\pm 0.32}$	$78.82{\scriptstyle\pm 2.53}$
w Rep. Clu. Init	${\color[rgb]{0.0781,0.5,0.332}\underline{\mathbf{87.87{\scriptstyle\pm 0.53}}}}$	$91.45{\scriptstyle\pm 0.23}$	$74.51{\scriptstyle\pm 0.95}$	$69.59{\scriptstyle\pm 1.18}$	$68.39{\scriptstyle\pm 1.09}$	$60.42{\scriptstyle\pm 13.45}$	$52.57{\scriptstyle\pm 1.31}$	$72.26{\scriptstyle\pm 1.70}$

TABLE IV: Performance of ARC zero and its variants in terms of AUROC.

Panel of Figure 9 — Fig. 5: Performance with varying $n_{k}$ in few-shot method ARC. The shaded areas represent the standard deviation over 5 repeated runs.

Panel of Figure 10 — Fig. 6: Performance with varying $n_{k}$ in zero-shot method ARC zero . The shaded areas represent the standard deviation over 5 repeated runs.

Fig. 7: Efficiency comparison of inference and fine-tuning time.

Panel of Figure 12 — Fig. 8: Visualization of the attention weights of cross-attentive scoring module. Darker colors represent larger values.

Panel of Figure 13 — Fig. 9: t-SNE visualization of representations by ARC and baselines.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 14 — Fig. 10: Hyperparameter sensitivity analysis.

Supervised - Pre-Train Only
Method	Cora	CiteSeer	ACM	BlogCatalog	Facebook	Weibo	Reddit
GCN	$7.41{\scriptstyle\pm 1.55}$	$6.40{\scriptstyle\pm 1.40}$	$5.27{\scriptstyle\pm 1.12}$	$7.44{\scriptstyle\pm 1.07}$	$1.59{\scriptstyle\pm 0.11}$	${\color[rgb]{0.9258,0.3281,0.0781}\underline{67.21{\scriptstyle\pm 15.20}}}$	$3.39{\scriptstyle\pm 0.39}$
GAT	$6.49{\scriptstyle\pm 0.84}$	$5.58{\scriptstyle\pm 0.62}$	$4.70{\scriptstyle\pm 0.75}$	$12.81{\scriptstyle\pm 2.08}$	$3.14{\scriptstyle\pm 0.37}$	$33.34{\scriptstyle\pm 9.80}$	$3.73{\scriptstyle\pm 0.54}$
BGNN	$4.90{\scriptstyle\pm 1.27}$	$3.91{\scriptstyle\pm 1.01}$	$3.48{\scriptstyle\pm 1.33}$	$5.73{\scriptstyle\pm 1.47}$	$3.81{\scriptstyle\pm 2.12}$	$30.26{\scriptstyle\pm 29.98}$	$3.52{\scriptstyle\pm 0.50}$
BWGNN	$7.25{\scriptstyle\pm 0.80}$	$6.35{\scriptstyle\pm 0.73}$	$7.14{\scriptstyle\pm 0.20}$	$8.99{\scriptstyle\pm 1.12}$	$2.54{\scriptstyle\pm 0.63}$	$12.13{\scriptstyle\pm 0.71}$	$3.69{\scriptstyle\pm 0.81}$
GHRN	$9.56{\scriptstyle\pm 2.40}$	$7.79{\scriptstyle\pm 2.01}$	$5.61{\scriptstyle\pm 0.71}$	$10.94{\scriptstyle\pm 2.56}$	$2.41{\scriptstyle\pm 0.62}$	$28.53{\scriptstyle\pm 7.38}$	$3.24{\scriptstyle\pm 0.33}$

Appendix

TABLE V: Anomaly detection performance in terms of AUPRC (in percent, mean

\pm

<

0.05).

Panel of Figure 16 — Fig. 11: Additional results for varying $n_{k}$ in the few-shot setting (ARC).

Panel of Figure 17 — Fig. 12: Additional results for varying $n_{k}$ in the zero-shot setting (ARC zero ).

Anomaly Detection

Anomaly detection finds rare or unexpected patterns in data, from defective products and system faults to fraud and security threats.

See all 629 papers

Jun 10, 2026

A Zero-shot Generalized Graph Anomaly Detection Framework via Node Reconstruction

Cross-domain graph anomaly detection (GAD) aims to identify abnormal nodes in unseen target graphs, showing strong potential in real-world applications with heterogeneous graph data. However, existing methods often depend on dataset-specific feature semantics and structural patterns, which limits their ability to generalize across different domains. To address this challenge, we propose AlignGAD, a zero-shot generalized graph anomaly detection framework. Our framework is built upon three key components: a Global Unification Module that aligns heterogeneous node features and normalizes graph signals in the spectral domain; a Clustering Module that constructs cluster-aware graph views to capture group-level abnormal patterns; and a Node Discrepancy Scoring Module that measures reconstruction discrepancy and aggregates anomaly evidence from different graph views. Experiments on multiple real-world datasets demonstrate the effectiveness of AlignGAD under the zero-shot GAD setting.

Phan Nguyen, Dat Cao, Hien Chu +1

School of Computing, KAIST

Cross-Domain Generalization Reconstruction-Based Anomaly Detection Zero-Shot Anomaly Detection 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

May 26, 2026

Detect by Yourself: Self-Designing Agentic Workflows for Few-Shot Graph Anomaly Detection

Graph anomaly detection aims to identify anomaly nodes in attributed graphs and plays an important role in real-world applications. However, existing graph anomaly detection methods still face two key challenges: 1) fixed pipelines, which restrict their adaptability across different graph tasks under limited supervision; 2) weak evidence, which prevents them from explicitly incorporating contextual and structural anomaly signals into the detection process. In this paper, we propose a novel framework, self-designing agentic workflows for few-shot graph anomaly detection (SignGAD). Specifically, we propose a novel paradigm that reformulates graph anomaly detection task from training a fixed anomaly detector to designing task-conditioned detection workflows. By constructing detection workflows, SignGAD selects suitable graph encodings and detector designs to exploit task-specific anomaly evidence. Meanwhile, we introduce a guarded final refit strategy to refine the selected workflow by calibrating refit acceptance, enhancing reliability under limited supervision. Extensive experiments conducted on several real-world datasets demonstrate that SignGAD achieves strong performance against state-of-the-art methods, highlighting its effectiveness on graph anomaly detection tasks.

Tairan Huang, Qiang Chen, Yili Wang +4

Central South University · The Hong Kong University of Science and Technology · The Hong Kong University of Science and Technology (Guangzhou)

Agentic Workflows Anomaly Detection

Ask PDF

Abstract

Anomaly Detection

Explore similar work

A Zero-shot Generalized Graph Anomaly Detection Framework via Node Reconstruction

Generalist Graph Anomaly Detection via Prototype-Based Distillation

Detect by Yourself: Self-Designing Agentic Workflows for Few-Shot Graph Anomaly Detection