One Pretrained Decision Model Is Enough for Training-Free Generalist Graph Anomaly Detection

Xinyu Zhao; Qingyun Sun; Ziwei Zhang; Tianchen Zhu; Daqing Li; Jianxin Li

cs.LGOct 22, 2026

One Pretrained Decision Model Is Enough for Training-Free Generalist Graph Anomaly Detection

Authors: Xinyu Zhao, Qingyun Sun, Ziwei Zhang, Tianchen Zhu, Daqing Li, Jianxin Li

Organizations: School of Computer Science and Engineering, Beihang University, Beijing, China · School of Reliability and Systems Engineering, Beihang University, Beijing, China

Abstract

Graph anomaly detection (GAD) aims to identify irregular entities in graph-structured data, yet heterogeneous anomaly patterns and distributional differences across graphs challenge generalizable detection. Existing training-free approaches avoid task-specific optimization but primarily rely on predefined numerical scoring rules, leaving the semantic interpretation of heterogeneous anomaly evidence largely unexplored. In this paper, we propose Jev-GAD, a training-free generalist graph anomaly detection framework that bridges heterogeneous numerical evidence and pretrained decision mechanisms. To address heterogeneous anomaly patterns, we construct complementary structural, attribute, and neighborhood evidence and transform heterogeneous statistics into graph-relative anomaly percentiles. To move beyond predefined numerical scoring rules, we leverage a single frozen pretrained language model for evidence-conditioned anomaly assessment. To bridge the semantic gap between numerical evidence and pretrained models, we design a semantic Graph-to-State (G2S) transformation that organizes graph context and semantically labeled node evidence into structured decision states for non-generative option-logit readout. The resulting decision scores are integrated with four complementary statistical scores through equal-weight rank-based fusion, without GAD-specific gradient-based optimization. Extensive experiments on seven graph anomaly detection benchmarks spanning citation, social, and co-review networks demonstrate the effectiveness of Jev-GAD, achieving average AUROC and AUPRC scores of 89.18% and 52.40%, respectively. Further analyses validate the contributions of pretrained decision readout and semantic evidence grounding.

Figures & tables

Figure 1. Motivation of Jev-GAD: heterogeneous anomalies, fixed scoring rules, and missing semantic grounding.

Figure 2. Overview of Jev-GAD. Heterogeneous anomaly evidence is organized into semantically grounded decision states for a single frozen pretrained model. Complementary statistical scoring and rank-based fusion produce the final anomaly ranking without GAD-specific gradient-based optimization.

Supervised Methods
Method (Venue)	ACM	Amazon	BlogCatalog	CiteSeer	Cora	Facebook	Weibo	Avg.
GCN ( ICLR’17 )	47.69 $\pm$ 1.48	41.10 $\pm$ 4.24	43.83 $\pm$ 2.39	53.83 $\pm$ 4.01	51.99 $\pm$ 3.20	46.09 $\pm$ 18.88	37.64 $\pm$ 13.84	46.02
GAT ( ICLR’18 )	46.54 $\pm$ 4.71	55.35 $\pm$ 5.69	49.32 $\pm$ 1.70	44.58 $\pm$ 8.25	44.65 $\pm$ 6.20	53.96 $\pm$ 26.14	35.50 $\pm$ 6.91	47.13
BWGNN ( ICML’22 )	67.43 $\pm$ 2.48	50.01 $\pm$ 0.04	69.91 $\pm$ 4.66	42.58 $\pm$ 2.86	39.75 $\pm$ 3.16	64.84 $\pm$ 4.17	50.08 $\pm$ 0.03	54.94
GHRN ( WWW’23 )	52.72 $\pm$ 1.01	63.52 $\pm$ 1.92	62.28 $\pm$ 2.42	54.11 $\pm$ 0.88	47.65 $\pm$ 2.03	54.04 $\pm$ 3.81	17.79 $\pm$ 5.00	50.30
Unsupervised Methods

Table 1. AUROC comparison on seven graph anomaly detection benchmarks (%). Results are reported as mean

\pm

standard deviation. Best and second-best mean values are highlighted in bold and underline , respectively.

Supervised Methods
Method (Venue)	ACM	Amazon	BlogCatalog	CiteSeer	Cora	Facebook	Weibo	Avg.
GCN ( ICLR’17 )	3.68 $\pm$ 0.16	5.54 $\pm$ 0.51	5.71 $\pm$ 0.38	5.21 $\pm$ 0.89	5.85 $\pm$ 0.55	4.55 $\pm$ 3.06	25.97 $\pm$ 15.84	8.07
GAT ( ICLR’18 )	3.38 $\pm$ 0.45	8.10 $\pm$ 1.24	6.64 $\pm$ 0.71	4.35 $\pm$ 0.82	4.93 $\pm$ 0.75	6.18 $\pm$ 5.83	15.86 $\pm$ 5.83	7.06
BWGNN ( ICML’22 )	19.48 $\pm$ 2.20	6.79 $\pm$ 0.03	28.37 $\pm$ 4.31	4.02 $\pm$ 0.43	5.30 $\pm$ 0.87	5.71 $\pm$ 1.60	10.47 $\pm$ 0.05	11.45
GHRN ( WWW’23 )	9.21 $\pm$ 2.94	11.18 $\pm$ 1.10	24.57 $\pm$ 3.60	5.37 $\pm$ 0.14	5.23 $\pm$ 0.19	3.61 $\pm$ 0.68	8.35 $\pm$ 2.44	9.65
Unsupervised Methods

Table 2. AUPRC comparison on seven graph anomaly detection benchmarks (%). Results are reported as mean

\pm

standard deviation. Best and second-best mean values are highlighted in bold and underline , respectively.

Figure 3. Diagnostic analyses on four graphs. (a) Component ablation. (b, c) Semantic grounding under Decision-only and full Jev-GAD scoring; Mismatched Semantics averages three fixed cyclic shifts. (d) Raw, aligned, and combined evidence views. (e, f) One-for-one replacement of statistical scorers. Y-axes do not all start at zero; numbers above bars are exact scores (%). Six grouped bar charts of AUROC and AUPRC on Amazon, BlogCatalog, CiteSeer, and Cora: component ablation; semantic grounding under decision-only and full Jev-GAD scoring; raw, aligned, and combined evidence views; and one-for-one statistical backend replacements on two pairs of graphs.

Figure 4. Prompt robustness. The first row of each block gives the original score (%); the other rows give changes relative to it in percentage points.

Figure 5. Sensitivity to decision weight

\lambda

on four graphs. The dashed line marks equal-weight fusion (

\lambda=0.2

), while shaded bands indicate

\pm 1

standard deviation across five Isolation Forest seeds. The endpoints represent Statistical-only (

\lambda=0

) and Decision-only (

\lambda=1

Figure 6. Computational efficiency. Numbers above bars are measured total times (s). Stacked bars of CPU processing time and frozen Qwen3-4B inference time on four graphs, with measured totals above the bars and node counts below.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Table 3 — Table 3. Node-level evidence fields in the semantic decision state.

Dataset	Domain	#Node	#Edge	#Feat.	#Anom.	Ratio	Deg.
ACM	Citation	16,484	74,073	8,337	597	3.62	8.99
CiteSeer	Citation	3,327	5,077	3,703	150	4.51	3.05
Cora	Citation	2,708	5,803	1,433	150	5.54	4.29
BlogCatalog	Social	5,196	172,783	8,189	300	5.77	66.51
Facebook	Social	1,081	27,552	576	25	2.31	50.98
Weibo	Social	8,405	377,271	400	868	10.33	89.77

Appendix

Table 4. Statistics of the seven benchmark graphs. Edges are undirected and self-loop-free; Ratio is the percentage of anomalous nodes and Deg. the average degree.

Dataset	5D	7D	10D
Amazon	37.48	41.26	42.34
BlogCatalog	32.86	33.50	33.73
CiteSeer	85.29	88.02	86.23
Cora	74.12	76.10	73.58
Macro average	57.44	59.72	58.97

Appendix

Table 5. AUPRC (%) of the fused method with different semantic state sizes. The 7D configuration is the default Jev-GAD setting.

Figure 7. Cumulative anomaly hits at different ranking depths on CiteSeer. Jev-GAD retrieves 122 true anomalies at Top-150, compared with 105 for Statistical-only. Cumulative anomaly hits across ranking depths on CiteSeer, comparing Jev-GAD with Statistical-only. At Top-150, Jev-GAD retrieves 122 anomalies, compared with 105 for Statistical-only.

Dataset	Stat.	Mean	Max	LOF	Semantic
Amazon	34.67	28.11	25.80	30.93	41.26
BlogCatalog	32.59	30.23	33.80	28.87	33.50
CiteSeer	71.06	70.29	71.36	56.13	88.02
Cora	67.34	71.29	69.20	52.63	76.10
Macro Avg.	51.41	49.98	50.04	42.14	59.72

Appendix

Table 6. Alternative fifth-scorer controls on four diagnostic datasets (AUPRC, %).

Dataset	Stat.	Mean	Max	LOF	Semantic
Amazon	0.592	0.353	0.314	0.511	0.611
BlogCatalog	0.613	0.495	0.566	0.544	0.593
CiteSeer	0.783	0.660	0.703	0.491	0.914
Cora	0.750	0.747	0.679	0.512	0.813
Macro Avg.	0.684	0.564	0.565	0.514	0.733

Appendix

Table 7. Pair-weighted within-bin AUC conditioned on Statistical-only anomaly scores.

Graph-Based Anomaly Detection

See all 38 papers

AnomalyGFM: Graph Foundation Model for Zero/Few-shot Anomaly Detection
Feb 13, 2025Hezhe Qiao, Chaoxi Niu, Ling Chen +1Graph-Based Anomaly Detection Anomaly Detection
School of Computing and Information Systems Singapore Management University Singapore · Australian Artificial Intelligence Institute University of Technology Sydney Sydney, Australia
Generalist Graph Anomaly Detection via Prototype-Based Distillation
May 26, 2026Yiming Xu, Zihan Chen, Zhen Peng +4Unsupervised Anomaly Detection Zero-Shot Anomaly Detection
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
FoundAna: A GNN-assisted Foundation Model for Graph Anomaly Detection
Sep 16, 2026Suprim Nakarmi, Chahana Dahal, Yue Zhao +2Graph Neural Networks Reconstruction-Based Anomaly Detection
Department of Computer Science, University of Nevada Las Vegas, Las Vegas, USA · Department of Computer Science, University of Southern California, Los Angeles, USA

Feb 13, 2025J/K move · Enter open · S save

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

Hezhe Qiao, Chaoxi Niu, Ling Chen, Guansong Pang

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

Graph-Based Anomaly Detection Anomaly Detection Zero-Shot Anomaly Detection Few-Shot Anomaly Detection Graph Foundation Models

Open paper AskPDF arXiv

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.

Abstract

Graph-Based Anomaly Detection

Explore similar work

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

Generalist Graph Anomaly Detection via Prototype-Based Distillation

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