How to Use Graph Data in the Wild to Help Graph Anomaly Detection?

Carl Yang; Chen Zhao; Chunping Wang; Jiaan Wang; Jiarong Xu; Yang Yang; Yuxuan Cao

REVIEW 5 major objections 6 minor 63 references

How to Use Graph Data in the Wild to Help Graph Anomaly Detection?

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Wild-GAD claims that training a target-initialized one-class model on a few external graphs chosen for representativity and diversity lifts graph anomaly detection, reporting average gains of 18% in AUCROC and 32% in AUCPR.

desk verdict A genuinely useful empirical framework for borrowing wild graph data in unsupervised GAD, with real but somewhat overstated headline gains and an unverified semantic-alignment assumption that deserves a closer look. read the letter →

arxiv 2506.04190 v1 pith:4F6QNPV4 submitted 2025-06-04 cs.LG

Yuxuan Cao , Jiarong Xu , Chen Zhao , Jiaan Wang , Carl Yang , Chunping Wang , Yang Yang This is my paper

classification cs.LG

keywords graphanomalydetectionexternaldataunsupervisedlearningone-classclassificationselectionhypersphericalenergyfeaturespacealignmentneuralnetworks

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that unsupervised graph anomaly detection, which must spot unusual nodes without any anomaly labels, can be improved by borrowing unlabeled graphs from other domains. The proposed framework, Wild-GAD, scores each candidate external graph by two criteria — representativity, or how well it matches the normal behavior the target graph has already learned, and diversity, or how much new normal variety it adds — and then continues training a target-initialized one-class model on the best-scoring external graphs. Across six real-world datasets the authors report average gains of 18% in AUCROC and 32% in AUCPR over the best-competing unsupervised baselines, and they report surpassing semi-supervised methods that are given up to 70% of the anomaly labels. The message a sympathetic reader takes away is that normal patterns are portable across graphs: with the right selection, plentiful unlabeled external data can supply the knowledge that scarce labels cannot.

What carries the argument

The carrying object is a spherical-coordinate view of node embeddings in the target's learned one-class space, plus the selection score built on it. Each embedding $z_i$ is written as $\rho_i = [r_i, \varphi_i]$ with radial distance $r_i = \|z_i - c\|$ from the sphere center $c$ (the mean embedding) and relative direction $\varphi_i = (z_i - c)/\|z_i - c\|$. Representativity combines a sphere-center restriction (Euclidean distance between the target-only and the mixed embedding centers) with a distributional-similarity term (Wasserstein distance between the spherical-coordinate distributions). Diversity is the minimum hyperspherical energy, motivated by the Thomson problem of spreading points evenly on a sphere: under a Gaussian potential kernel $e(u,v) = e^{-t\|u-v\|^2}$, close pairs of embeddings contribute more energy, so lower total energy means the external points are scattered rather than clustered and thereby cover more of the normal-pattern sphere across radii. The three $z$-normalized terms are merged into $J(G) = \eta\,\mathrm{MEAN}(\hat{s}^c_{\mathrm{sim}} + \hat{s}^d_{\mathrm{sim}}) + (1-\eta)\,\hat{s}_{\mathrm{div}}$, and the $k$ graphs with the smallest scores are selected. Training then proceeds target-first: the one-class model is trained on the target graph, then continued on the selected external graphs with a Fisher-information regularizer, so external knowledge enlarges the normal region without erasing what the target itself taught.

What would settle it

On a held-out target graph, replace the representativity-diversity selection with the same number of external graphs drawn uniformly at random from UniWildGraph over many seeds: if random graphs match or beat the selected ones in AUCROC and AUCPR, the selection score is not the cause of the reported gains. To test the feature-alignment premise directly, add external graphs from domains far outside the four built into UniWildGraph, such as molecular or protein graphs, through the same mBERT textualization and rerun the pipeline: if the average 18% AUCROC and 32% AUCPR gains disappear, the unified semantic space holds only within the curated domains and the selection scores lose their grounding.

Watch

Extended reading notes

Core claim

Wild-GAD claims that external graph data can serve as a reliable source of normal-pattern knowledge for graph anomaly detection. The framework first builds a candidate pool, UniWildGraph, from twelve raw graphs spanning e-commerce, social, citation, and hyper-link domains, augments each graph by feature masking, node dropping, edge perturbation, and subgraph extraction at four ratios (221 graphs total), and then textualizes every node's features and maps them into one semantic space with the multilingual language model mBERT. For a target graph, a base one-class model (Deep SVDD) is trained on the target alone, and the resulting embedding sphere is used to score each external graph by how little it shifts the sphere center and by the Wasserstein distance between target and mixed spherical-coordinate distributions (representativity), and by how uniformly it spreads points on the hypersphere via minimum hyperspherical energy (diversity). The top-scoring graphs are then used to continue training the target-initialized model under a Fisher-information penalty that keeps new parameters close to the target-trained ones. On six real-world datasets the paper reports average improvements of 18% in AUCROC and 32% in AUCPR over the best-competing unsupervised methods, with ablations showing that both selection criteria matter; an extension using labeled external anomaly data helps in some cases but can also hurt when the external abnormal patterns differ from the target's, a limitation the paper itself notes.

Load-bearing premise

The load-bearing premise is that textualizing all node features and embedding them with one multilingual language model puts nodes from different domains in a genuinely shared semantic space, so that distances between target nodes and external nodes in the learned sphere are meaningful; if the alignment is not semantically faithful, the representativity and diversity scores are computed in arbitrary coordinates and the selection step has no grounding.

Editorial extensions

If this is right

Any target graph, however small or label-poor, can be paired with external graphs from a shared pool, so unsupervised graph anomaly detection no longer depends only on the normal patterns visible in the target graph itself.
The selection score doubles as a predictor of data quality: the paper reports Pearson correlations between the estimated score and downstream AUCROC in the range 0.52–0.77, so the same score can be used to decide whether external data will help a given target.
Adding more external graphs generally helps or is neutral: moving from one to two selected graphs improves most datasets, and further increases to three or five hold performance steady, so over-selection is not a serious risk.
The benefit transfers across backbone models: both the one-class SVDD variant and the graph-autoencoder variant improve, so the external-data recipe is not tied to a single anomaly-scoring mechanism.
On YelpHotel, the framework surpasses semi-supervised GNN baselines at label ratios up to 70%, indicating that unlabeled external graphs can substitute for a large fraction of manual anomaly labels.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable extension the paper leaves implicit: if the mechanism is really 'external data fills in missing normal patterns,' the gains should concentrate on small or skewed target graphs and shrink toward zero on large, self-sufficient ones; this could be checked by ablating target graph size while holding the candidate pool fixed.
The paper never varies the training order (target first, then external graphs), so it is an open question whether the Fisher-information penalty is doing the forgetting-prevention work or whether the order itself matters more; training jointly on mixed data or in the reverse order would separate the two.
The mBERT alignment is the unstated dependency of the whole pipeline; swapping the language model or the tabular-to-text rule and rechecking whether the selection score still predicts performance would reveal how much of the gain comes from alignment versus from the selection-plus-continued-training recipe itself.
For practitioners, the reported score–performance correlation makes the selection score usable as a free go/no-go diagnostic: score candidate external graphs first, and only invest in continued training when the top-scoring graphs are also semantically plausible matches for the target.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

This is the first serious attempt I know of to use unlabeled external graphs for unsupervised graph anomaly detection, and the basic recipe works. You build a large external graph database with a unified feature space, select graphs by representativity and diversity in the target's embedding space, then continue training the target-initialized model with an EWC-style regularizer. The reported gains over strong baselines are real on six datasets, and the ablations show both selection criteria matter. The correlation between selection score and downstream AUC is a nice sanity check, though it is internal to the same mBERT-based pipeline. The authors also ship code and are unusually candid in the appendices about noise from external data and the relationship to pre-training/fine-tuning. That's solid work.

The soft spots are real but mostly fixable. The abstract's "18% AUCROC and 32% AUCPR improvement" is never precisely defined; it appears to be relative improvement over the best competing method per dataset, but the tables don't easily reproduce those numbers, and the text doesn't say whether it's absolute or relative. The semi-supervised claim ("up to 70% label ratio") is tested on YelpHotel only, yet is stated without that qualifier. Equation 5 defines diversity as an integral over probability measures on spheres, which is not an operationalized computation; how was it actually computed? Hyperparameter sensitivity (eta, t, tau, beta) is not reported.

The stress-test concern about the mBERT feature space is the most substantive issue. Everything hinges on the claim that textualized attributes from, say, a Tolokers worker profile and an arXiv abstract live in a common semantic space. If that alignment is not meaningful, the representativity and diversity scores are computed on arbitrary coordinates. The paper provides no external validation of this alignment, and the internal correlations could be driven by the shared encoder rather than by genuine semantic similarity. I don't think this kills the empirical claim, because the method still improves detection even if the selection criteria are imperfect signals, but it is a load-bearing assumption that should be probed with a different feature encoder or a semantic ground-truth test.

This paper deserves a serious referee and a revision. The empirical contribution is real, the database is a resource, and the limitations are addressable. Practitioners in fraud and bot detection will use this; researchers will argue about the feature-space question. I would accept it with major revisions that (1) define the headline gains precisely, (2) operationalize Eq. 5, (3) report sensitivity, and (4) either validate or explicitly qualify the semantic alignment assumption.

Referee Report

5 major / 6 minor

Summary. The paper proposes Wild-GAD, a framework for unsupervised graph anomaly detection that leverages external unlabeled graph data. It constructs UniWildGraph, a database of twelve raw graphs across four domains plus augmented variants (221 graphs total), textualizes node attributes and embeds them with mBERT to create a supposedly unified feature space. A selection strategy scores candidate external graphs by representativity (center shift and Wasserstein distance) and diversity (hyperspherical energy), and the selected graphs are used to continue training a target-initialized one-class SVDD or graph autoencoder with an EWC-style parameter regularizer. Experiments on six real-world anomaly detection datasets compare Wild-GAD against unsupervised baselines and against semi-supervised methods at varying label ratios.

Significance. If the central claims hold, the paper would make a useful contribution to graph anomaly detection by showing that unlabeled external graphs can serve as a source of normal-pattern knowledge, alleviating label scarcity. The manuscript has concrete strengths: it assembles a large and diverse candidate database; it evaluates on six real-world datasets with organic anomalies; it provides ablations showing that both selection criteria contribute (Tables 5 and 6); it reports correlations between the proposed score and downstream performance (Figures 2 and 3); and it releases code. However, several load-bearing aspects need tightening before the claims are fully supported: the operational definition of the diversity measure, the reproducibility of the headline average gains from the tables, the validation of the unified feature-space assumption, the underspecified labeled-data extension, and the overbroad semi-supervised comparison claim.

major comments (5)

[Abstract and Section 5.2] The abstract and Section 5.2 state that Wild-GAD achieves an average 18% AUCROC and 32% AUCPR improvement over the best-competing methods. I could not reproduce these numbers from Tables 3 and 4. For the strongest unsupervised variant, Wild-GAD(OC/M/UL), the per-dataset relative AUCROC improvements over the best baseline are approximately 9.7%, 5.9%, 17.1%, 7.4%, -0.3%, and 10.2%, averaging about 8.3%, not 18%. For AUCPR the relative improvements average about 50%, largely driven by AmazonCN, not 32%. Please clarify how the reported averages were computed, which variant and which baseline set they use, and ensure consistency between the abstract and the tables.
[Section 4.2, Eq. (5)] The diversity measure is defined as a continuous integral over probability measures on concentric spheres, but the paper does not specify how this integral is computed in practice. Eq. (5) involves integrals over u, v, and r with u, v drawn from M(S(r)), yet no discretization, sampling scheme, or radius-shell procedure is given. Furthermore, the Gaussian kernel in Eq. (6) depends on a hyperparameter t whose value is not reported. Since the diversity score is a load-bearing part of the selection objective in Eq. (7), the paper must provide a reproducible computational recipe and sensitivity analysis for this component.
[Section 4.1 and 4.2] The framework's load-bearing assumption is that mBERT-textualized node attributes from heterogeneous domains occupy a unified semantic space, so that spherical coordinates computed from target and external embeddings in Section 4.2 are meaningfully comparable. The paper provides no direct validation of this alignment. The correlation evidence in Figures 2 and 3 is internal in the sense that both the selection score and the final performance are produced by the same mBERT-based pipeline. A concrete test is needed: for example, compare mBERT-based cross-domain similarities against some semantic ground truth, or replace mBERT with an alternative feature encoder and check whether the selection mechanism still transfers. Without such evidence, the reported gains could be artifacts of the specific embedding rather than properties of the external data.
[Section 4.4 and Tables 3-4] The labeled external data variants (Wild-GAD .../L) are presented as main results, but Section 4.4 does not specify how labels are used in the training loss of Eq. (8). The text says that selection scores are computed separately for normal and abnormal samples, but it is not explained whether labeled abnormal samples are excluded from the one-class training set, whether a supervised loss replaces L_ano, or how the extension interacts with the regularizer. Without a precise protocol, the L variants cannot be reproduced or properly interpreted.
[Section 5.2 and Figure 4] The claim that Wild-GAD outperforms semi-supervised methods at label ratios up to 70% is supported only by Figure 4, whose caption states that the experiments are on YelpHotel. The abstract and Section 5.2 state this as a general result without qualification. Please either provide results on additional datasets or explicitly restate the claim as being specific to YelpHotel.

minor comments (6)

[Eq. (4)] There is a typo: 'Warsenstein' should be 'Wasserstein'.
[Figure 1] The label 'Instgram' should be 'Instagram'.
[Appendix C] The appendix states that the application scope can be characterized by a threshold on the calculated data score, but no threshold is ever defined or applied in the experiments. Please clarify or remove this claim.
[Section 4.2] The sentence 'As previously discussed in section 2' appears to reference the wrong section; the selection discussion is in Section 4.2 and the relevant background is in Section 1 or 2.
[Eq. (4)] The notation P(\rho_in) and P(\rho_mix) is used without defining the distribution; please clarify whether these are empirical distributions over the corresponding sets of spherical coordinates.
[Table 2] The Enron dataset has only five abnormal nodes (0.04% abnormal ratio); the stability of AUCPR and AUROC comparisons on this dataset should be discussed, especially because many reported standard deviations are large relative to the mean differences.

Circularity Check

0 steps flagged · score 2.0 of 10

No definitional circularity; only minor, non-load-bearing self-citations.

full rationale

No step in the paper reduces a stated prediction to an input by construction. The selection score in Eq. 7 is computed from embeddings produced by the target-trained model phi_0, and Eq. 8 initializes continued training from phi_0's parameters; this is a self-referential loop, but it is not a circular reduction. The selection objective never uses anomaly labels or test-set outcomes, and the reported AUC-ROC and AUC-PR gains are measured on held-out anomaly labels after training on the selected external graphs. The Pearson correlations in Figures 2 and 3 are empirical associations between selection scores and downstream performance, not identities. Self-citations [5, 14, 24, 40, 41, 51] appear only in related-work and fine-tuning contrasts, and the textualization rule from [41] is a simple, non-load-bearing implementation detail. No uniqueness theorem, imported ansatz, or renaming of a known result is used. The framework's central claim is therefore supported by an externally evaluated experiment rather than by definition.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim depends on hand-set hyperparameters (eta, beta, t, augmentation ratios) and on domain assumptions about a unified feature space and one-class normality. No target anomaly labels are used in selection, which keeps the empirical evaluation independent, but the diversity measure in Eq. 5 is not operationalized in the text.

free parameters (4)

eta = 0.5
Trade-off between representativity and diversity in Eq. 7; fixed without sensitivity analysis.
beta = 0.5
One-class SVDD error and support bound in Eq. 1; set globally even on datasets with more than 35% anomalies.
Gaussian kernel t
Width of the Gaussian potential kernel in Eq. 6; no value or sensitivity analysis is reported.
augmentation ratios tau = {0.2, 0.4, 0.6, 0.8}
Hand-chosen ratios used to construct the 221-graph candidate database in Section 4.1.

assumptions (4)

domain assumption One-class SVDD hypersphere objective yields a valid unsupervised anomaly score for all six test graphs
Adopted from Deep SVDD in Eq. 1; assumes anomalies are far from the center and are a minority, an assumption violated on C-15 and Twitter-20 which have 35.5% and 56.35% anomalies.
domain assumption mBERT-textualized attributes form a unified semantic space across domains
Section 4.1 'Feature Space Alignment'; all cross-graph comparisons in Section 4.2 rely on this alignment.
ad hoc to paper Minimizing hyperspherical energy measures the diversity relevant to normal-pattern coverage
Eqs. 5 and 6 are motivated by the Thomson problem, but no operational algorithm or justification specific to anomaly detection is provided.
domain assumption EWC-style Fisher information regularization prevents catastrophic forgetting of target knowledge
Eq. 8 uses a Fisher information matrix F, but the paper does not describe how F is estimated for the graph one-class model; if F is inaccurate, the target knowledge could be overwritten.

how reviews work

0 comments

Cite this review

Pith. "Pith review of How to Use Graph Data in the Wild to Help Graph Anomaly Detection?." pith.science (2026). https://pith.science/paper/4F6QNPV4

@misc{pith2026250604190,
  author       = {Pith},
  title        = {Pith review of: How to Use Graph Data in the Wild to Help Graph Anomaly Detection?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4F6QNPV4}},
  note         = {Machine review of arXiv:2506.04190}
}

read the original abstract

In recent years, graph anomaly detection has found extensive applications in various domains such as social, financial, and communication networks. However, anomalies in graph-structured data present unique challenges, including label scarcity, ill-defined anomalies, and varying anomaly types, making supervised or semi-supervised methods unreliable. Researchers often adopt unsupervised approaches to address these challenges, assuming that anomalies deviate significantly from the normal data distribution. Yet, when the available data is insufficient, capturing the normal distribution accurately and comprehensively becomes difficult. To overcome this limitation, we propose to utilize external graph data (i.e., graph data in the wild) to help anomaly detection tasks. This naturally raises the question: How can we use external data to help graph anomaly detection tasks? To answer this question, we propose a framework called Wild-GAD. It is built upon a unified database, UniWildGraph, which comprises a large and diverse collection of graph data with broad domain coverage, ample data volume, and a unified feature space. Further, we develop selection criteria based on representativity and diversity to identify the most suitable external data for anomaly detection task. Extensive experiments on six real-world datasets demonstrate the effectiveness of Wild-GAD. Compared to the baseline methods, our framework has an average 18% AUCROC and 32% AUCPR improvement over the best-competing methods.

Figures

Figures reproduced from arXiv: 2506.04190 by the authors.

**Figure 3.** The score of external data vs. the detection perfor [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

**Figure 4.** Comparison with semi-supervised methods under varying [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 2.** Figure 2: The score of external data vs. the detection perfor [PITH_FULL_IMAGE:figures/full_fig_p009_2.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 56 canonical work pages

[1]

Philip Bachman, R Devon Hjelm, and William Buchwalter. 2019. Learning repre- sentations by maximizing mutual information across views. InNeurIPS, Vol. 32

work page 2019
[2]

Sambaran Bandyopadhyay, Lokesh N, Saley Vishal Vivek, and M Narasimha Murty. 2020. Outlier resistant unsupervised deep architectures for attributed network embedding. InWSDM. 25–33

work page 2020
[3]

Yuanchen Bei, Sheng Zhou, Qiaoyu Tan, Hao Xu, Hao Chen, Zhao Li, and Jiajun Bu. 2023. Reinforcement neighborhood selection for unsupervised graph anomaly detection. InICDM. 11–20

work page 2023
[4]

Liron Bergman, Niv Cohen, and Yedid Hoshen. 2020. Deep nearest neighbor anomaly detection.arXiv preprint arXiv:2002.10445(2020)

arXiv 2020
[5]

Yuxuan Cao, Jiarong Xu, Carl Yang, Jiaan Wang, Yunchao Zhang, Chunping Wang, Lei CHEN, and Yang Yang. 2023. When to Pre-Train Graph Neural Networks? From Data Generation Perspective!. InSIGKDD. 142–153

work page 2023
[6]

Ziwei Chai, Yang Yang, Jiawang Dan, Sheng Tian, Changhua Meng, Weiqiang Wang, and Yifei Sun. 2023. Towards learning to discover money laundering sub-network in massive transaction network. InAAAI

work page 2023
[7]

Ziwei Chai, Siqi You, Yang Yang, Shiliang Pu, Jiarong Xu, Haoyang Cai, and Weihao Jiang. 2022. Can abnormality be detected by graph neural networks?. In IJCAI. 1945–1951

work page 2022
[8]

Zhenxing Chen, Bo Liu, Meiqing Wang, Peng Dai, Jun Lv, and Liefeng Bo. 2020. Generative adversarial attributed network anomaly detection. InCIKM

work page 2020

Show all 63 references

[9]

Henry Cohn and Abhinav Kumar. 2007. Universally optimal distribution of points on spheres.Journal of the American Mathematical Society20, 1 (2007), 99–148

2007
[10]

Stefano Cresci, Roberto Di Pietro, Marinella Petrocchi, Angelo Spognardi, and Maurizio Tesconi. 2015. Fame for sale: Efficient detection of fake Twitter followers. Decision Support Systems80 (2015), 56–71

2015
[11]

Alex Davies and Nirav Ajmeri. 2022. Realistic Synthetic Social Networks with Graph Neural Networks. arXiv:2212.07843 [cs.SI]

2022 arXiv
[12]

Kaize Ding, Jundong Li, Rohit Bhanushali, and Huan Liu. 2019. Deep anomaly detection on attributed networks. InSIAM. 594–602

2019
[13]

Haoyi Fan, Fengbin Zhang, and Zuoyong Li. 2020. Anomalydae: Dual autoencoder for anomaly detection on attributed networks. InICASSP. IEEE, 5685–5689

2020
[14]

Taoran Fang, Yunchao Zhang, Yang Yang, Chunping Wang, and Lei Chen. 2024. Universal prompt tuning for graph neural networks. InNeurIPS, Vol. 36

2024
[15]

Shangbin Feng, Herun Wan, Ningnan Wang, Jundong Li, and Minnan Luo. 2021. Twibot-20: A comprehensive twitter bot detection benchmark. InCIKM

2021
[16]

Ralph Foorthuis. 2021. On the nature and types of anomalies: a review of devia- tions in data.International journal of data science and analytics(2021), 297–331

2021
[17]

Jintong Gao, He Zhao, Dan dan Guo, and Hongyuan Zha. 2024. Distribution alignment optimization through neural collapse for long-tailed classification. In ICML

2024
[18]

Shuzheng Gao, Wenxin Mao, Cuiyun Gao, Li Li, Xing Hu, Xin Xia, and Michael R. Lyu. 2024. Learning in the Wild: Towards Leveraging Unlabeled Data for Effec- tively Tuning Pre-Trained Code Models. InICSE. 969–981

2024
[19]

Yuan Gao, Xiang Wang, Xiangnan He, Zhenguang Liu, Huamin Feng, and Yong- dong Zhang. 2023. Addressing heterophily in graph anomaly detection: A per- spective of graph spectrum. InWWW. 1528–1538

2023
[20]

Hakim Hafidi, Mounir Ghogho, Philippe Ciblat, and Ananthram Swami. 2020. GraphCL: Contrastive self-supervised learning of graph representations.arXiv preprint arXiv:2007.08025(2020)

2020 arXiv
[21]

Xiaotian Han, Zhimeng Jiang, Ninghao Liu, and Xia Hu. 2022. G-Mixup: Graph Data Augmentation for Graph Classification. InICML

2022
[22]

Mingguo He, Zhewei Wei, Hongteng Xu, et al. 2021. Bernnet: Learning arbitrary graph spectral filters via bernstein approximation. InNeurIPS. 14239–14251

2021
[23]

Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. 2018. Deep anomaly detection with outlier exposure.arXiv preprint arXiv:1812.04606(2018)

2018 arXiv
[24]

Renhong Huang, Jiarong Xu, Xin Jiang, Chenglu Pan, Zhiming Yang, Chunping Wang, and Yang Yang. 2024. Measuring Task Similarity and Its Implication in Fine-Tuning Graph Neural Networks. InAAAI, Vol. 38. 12617–12625

2024
[25]

Xuanwen Huang, Yang Yang, Yang Wang, Chunping Wang, Zhisheng Zhang, Jiarong Xu, Lei Chen, and Michalis Vazirgiannis. 2022. Dgraph: A large-scale financial dataset for graph anomaly detection. InNeurIPS. 22765–22777

2022
[26]

Thomas N Kipf and Max Welling. 2016. Variational graph auto-encoders.arXiv preprint arXiv:1611.07308(2016)

2016 arXiv
[27]

Bryan Klimt and Yiming Yang. 2004. Introducing the Enron corpus. InCEAS. 1

2004
[28]

Rongmei Lin, Weiyang Liu, Zhen Liu, Chen Feng, Zhiding Yu, James M Rehg, Li Xiong, and Le Song. 2020. Regularizing neural networks via minimizing hyperspherical energy. InCVPR. 6917–6927

2020
[29]

Kay Liu, Yingtong Dou, Xueying Ding, Xiyang Hu, Ruitong Zhang, Hao Peng, Lichao Sun, and Philip S. Yu. 2024. PyGOD: A Python Library for Graph Outlier Detection.Journal of Machine Learning Research25, 141 (2024), 1–9

2024
[30]

Weiyang Liu, Rongmei Lin, Zhen Liu, Lixin Liu, Zhiding Yu, Bo Dai, and Le Song
[31]

Yixin Liu, Zhao Li, Shirui Pan, Chen Gong, Chuan Zhou, and George Karypis
[32]

Xiaoxiao Ma, Ruikun Li, Fanzhen Liu, Kaize Ding, Jian Yang, and Jia Wu. 2024. Graph anomaly detection with few labels: A data-centric approach. InSIGKDD. 2153–2164

2024
[33]

Xiaoxiao Ma, Jia Wu, Shan Xue, Jian Yang, Chuan Zhou, Quan Z Sheng, Hui Xiong, and Leman Akoglu. 2023. A Comprehensive Survey on Graph Anomaly Detection With Deep Learning.TKDE35, 12 (2023), 12012–12038

2023
[34]

Arjun Mukherjee, Vivek Venkataraman, Bing Liu, and Natalie Glance. 2013. What yelp fake review filter might be doing?. InICWSM, Vol. 7. 409–418

2013
[35]

Oleg Platonov, Denis Kuznedelev, Michael Diskin, Artem Babenko, and Liudmila Prokhorenkova. 2023. A critical look at the evaluation of GNNs under heterophily: Are we really making progress?arXiv preprint arXiv:2302.11640(2023)

2023 arXiv
[36]

Hezhe Qiao and Guansong Pang. 2024. Truncated affinity maximization: One- class homophily modeling for graph anomaly detection. InNeurIPS, Vol. 36

2024
[37]

Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. 2021. Panda: Adapting pretrained features for anomaly detection and segmentation. InCVPR. 2806–2814

2021
[38]

Tal Reiss and Yedid Hoshen. 2023. Mean-shifted contrastive loss for anomaly detection. InAAAI, Vol. 37. 2155–2162

2023
[39]

Lukas Ruff, Robert Vandermeulen, Nico Goernitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Alexander Binder, Emmanuel Müller, and Marius Kloft. 2018. Deep one-class classification. InICML. PMLR, 4393–4402

2018
[40]

Yifei Sun, Qi Zhu, Yang Yang, Chunping Wang, Tianyu Fan, Jiajun Zhu, and Lei Chen. 2024. Fine-Tuning Graph Neural Networks by Preserving Graph Generative Patterns. InAAAI. 9053–9061

2024
[41]

Yanchao Tan, Zihao Zhou, Hang Lv, Weiming Liu, and Carl Yang. 2024. Walklm: A uniform language model fine-tuning framework for attributed graph embedding. InNeurIPS, Vol. 36

2024
[42]

Jianheng Tang, Jiajin Li, Ziqi Gao, and Jia Li. 2022. Rethinking graph neural networks for anomaly detection. InInternational Conference on Machine Learning. PMLR, 21076–21089

2022
[43]

Joseph John Thomson. 1904. XXIV. On the structure of the atom: an investigation of the stability and periods of oscillation of a number of corpuscles arranged at equal intervals around the circumference of a circle; with application of the results to the theory of atomic struc...

1904
[44]

Leonid Nisonovich Vaserstein. 1969. Markov processes over denumerable prod- ucts of spaces, describing large systems of automata.Problemy Peredachi Infor- matsii5, 3 (1969), 64–72

1969
[45]

Huan Wang, Jia Wu, Wenbin Hu, and Xindong Wu. 2019. Detecting and assessing anomalous evolutionary behaviors of nodes in evolving social networks.TKDD 13, 1 (2019), 1–24

2019
[46]

Qizhou Wang, Guansong Pang, Mahsa Salehi, Wray Buntine, and Christopher Leckie. 2023. Cross-domain graph anomaly detection via anomaly-aware con- trastive alignment. InAAAI, Vol. 37. 4676–4684

2023
[47]

Xuhong Wang, Baihong Jin, Ying Du, Ping Cui, Yingshui Tan, and Yupu Yang
[48]

Yanling Wang, Jing Zhang, Shasha Guo, Hongzhi Yin, Cuiping Li, and Hong Chen. 2021. Decoupling representation learning and classification for gnn-based anomaly detection. InSIGIR. 1239–1248

2021
[49]

Shijie Wu and Mark Dredze. 2019. Beto, Bentz, Becas: The Surprising Cross- Lingual Effectiveness of BERT. InEMNLP. 833–844

2019
[50]

One-class graph neural networks for anomaly detection in attributed networks.Neural computing and applications33 (2021), 12073–12085

2021
[51]

Jiarong Xu, Renhong Huang, Xin Jiang, Yuxuan Cao, Carl Yang, Chunping Wang, and Yang Yang. 2023. Better with Less: A Data-Active Perspective on Pre-Training Graph Neural Networks. InNeurIPS, Vol. 36. 56946–56978

2023
[52]

Xiongxiao Xu, Kaize Ding, Canyu Chen, and Kai Shu. 2023. MetaGAD: Learn- ing to Meta Transfer for Few-shot Graph Anomaly Detection.arXiv preprint arXiv:2305.10668(2023)

2023 arXiv
[53]

Zhiang Wu, Guannan Liu, Junjie Wu, and Yong Tan. 2023. Are neighbors alike? A semisupervised probabilistic collaborative learning model for online review spammers detection.Information Systems Research(2023)

2023
[54]

Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. 2020. Graph contrastive learning with augmentations. InNeurIPS, Vol. 33. 5812–5823

2020
[55]

Bing Yu, Haoteng Yin, and Zhanxing Zhu. 2017. Spatio-temporal graph con- volutional networks: A deep learning framework for traffic forecasting.arXiv preprint arXiv:1709.04875(2017)

2017 arXiv
[56]

Zhiming Xu, Xiao Huang, Yue Zhao, Yushun Dong, and Jundong Li. 2022. Con- trastive attributed network anomaly detection with data augmentation. InPAKDD. 444–457

2022
[57]

Woodford, Meng Jiang, and Neil Shah

Tong Zhao, Yozen Liu, Leonardo Neves, Oliver J. Woodford, Meng Jiang, and Neil Shah. 2021. Data Augmentation for Graph Neural Networks. InAAAI

2021
[58]

Jiong Zhu, Yujun Yan, Lingxiao Zhao, Mark Heimann, Leman Akoglu, and Danai Koutra. 2020. Beyond homophily in graph neural networks: Current limitations and effective designs. InNeurIPS, Vol. 33. 7793–7804

2020
[59]

Shuke Zhang, Yanzhao Jin, Tianmeng Liu, Qi Wang, Zhaohui Zhang, Shuliang Zhao, and Bo Shan. 2022. SS-GNN: A Simple-Structured Graph Neural Network for Affinity Prediction. arXiv:2206.07015 [q-bio.BM]

2022 arXiv
[62]

data in the wild

Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang. 2021. Graph Contrastive Learning with Adaptive Augmentation. InWWW. How to Use Graph Data in the Wild to Help Graph Anomaly Detection? KDD ’25, August 3–7, 2025, Toronto, ON, Canada A Datasets Details In YelpH...

2021
[63]

F GAE Backbone Our framework is primarily designed with one-class SVDD as the backbone

This can be explained by the fact that, after a certain point, ad- ditional external data may not contribute significantly to learning new normal patterns, as the model may have already captured the essential behaviors from a smaller subset of the external data. F GAE Backbone...

2025
[2018]

InNeurIPS, Vol

Learning towards minimum hyperspherical energy. InNeurIPS, Vol. 31
[2021]

Anomaly detection on attributed networks via contrastive self-supervised learning.TNNLS33, 6 (2021), 2378–2392

2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Philip Bachman, R Devon Hjelm, and William Buchwalter. 2019. Learning repre- sentations by maximizing mutual information across views. InNeurIPS, Vol. 32

work page 2019

[2] [2]

Sambaran Bandyopadhyay, Lokesh N, Saley Vishal Vivek, and M Narasimha Murty. 2020. Outlier resistant unsupervised deep architectures for attributed network embedding. InWSDM. 25–33

work page 2020

[3] [3]

Yuanchen Bei, Sheng Zhou, Qiaoyu Tan, Hao Xu, Hao Chen, Zhao Li, and Jiajun Bu. 2023. Reinforcement neighborhood selection for unsupervised graph anomaly detection. InICDM. 11–20

work page 2023

[4] [4]

Liron Bergman, Niv Cohen, and Yedid Hoshen. 2020. Deep nearest neighbor anomaly detection.arXiv preprint arXiv:2002.10445(2020)

arXiv 2020

[5] [5]

Yuxuan Cao, Jiarong Xu, Carl Yang, Jiaan Wang, Yunchao Zhang, Chunping Wang, Lei CHEN, and Yang Yang. 2023. When to Pre-Train Graph Neural Networks? From Data Generation Perspective!. InSIGKDD. 142–153

work page 2023

[6] [6]

Ziwei Chai, Yang Yang, Jiawang Dan, Sheng Tian, Changhua Meng, Weiqiang Wang, and Yifei Sun. 2023. Towards learning to discover money laundering sub-network in massive transaction network. InAAAI

work page 2023

[7] [7]

Ziwei Chai, Siqi You, Yang Yang, Shiliang Pu, Jiarong Xu, Haoyang Cai, and Weihao Jiang. 2022. Can abnormality be detected by graph neural networks?. In IJCAI. 1945–1951

work page 2022

[8] [8]

Zhenxing Chen, Bo Liu, Meiqing Wang, Peng Dai, Jun Lv, and Liefeng Bo. 2020. Generative adversarial attributed network anomaly detection. InCIKM

work page 2020