CRoC: Context Refactoring Contrast for Graph Anomaly Detection with Limited Supervision

Da Sun Handason Tam; Siyue Xie; Wing Cheong Lau

REVIEW 3 major objections 3 minor 40 references

CRoC: Context Refactoring Contrast for Graph Anomaly Detection with Limited Supervision

T0 review · 3 major / 3 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Context Refactoring Contrast (CRoC) claims GNNs can detect graph anomalies with very few labels by recomposing each node's attributes while keeping its interactions, then using contrastive learning on unlabeled data, gaining up to 14% AUC…

desk verdict A plausible and genuinely distinct GAD idea that I cannot fully judge from this version because the full text is unreadable mojibake; worth chasing the clean PDF before deciding. read the letter →

arxiv 2508.12278 v2 pith:VFTR4GI3 submitted 2025-08-17 cs.LG cs.AI

Siyue Xie , Da Sun Handason Tam , Wing Cheong Lau This is my paper · ORCID

classification cs.LGcs.AI

keywords graphanomalydetectionlimitedsupervisioncontrastivelearningcontextrefactoringneuralnetworksheterogeneousrelationsnodeaugmentationrareclass

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 tries to establish that scarce labeled anomalies are not a hard bottleneck for graph anomaly detection. It introduces CRoC, which builds augmented graph views by recomposing node attributes within the existing interaction structure, then trains a GNN with a contrastive objective that uses both the few labeled nodes and the many unlabeled ones. Because anomalies are rare, most nodes are normal, so this attribute recomposition mostly creates realistic normal-context views rather than destroying the signal. The paper reports up to 14% AUC improvement over baseline GNNs and better results than existing graph anomaly detection methods on seven real-world datasets with limited labels. If correct, this gives practitioners a label-efficient way to identify rare suspicious nodes.

What carries the argument

The central mechanism is the context-refactoring augmentation: for each node, build its augmented feature by recomposing attributes from nodes that share its interaction context while leaving the edge structure unchanged, then use these augmented graphs as positive views in a contrastive loss. The mechanism is completed by encoding each relation type separately and fusing the relation-specific messages during propagation. It does two jobs: it generates training signal from unlabeled nodes by making the model invariant to non-semantic attribute variation, and it hardens the model against camouflage, where anomalies disguise themselves by imitating normal contexts.

What would settle it

Train CRoC and an identical model with the same contrastive loss but random attribute shuffling across nodes instead of context-aware refactoring; if the random-shuffle version matches or beats CRoC's AUC on the same datasets with the same label budget, then the claim that context refactoring preserves semantics and causes the improvement is not supported. A second check is to swap a fraction of positive training labels with normal labels and see whether CRoC's gains vanish more quickly than a supervised baseline's, which would indicate the method relies on label quality rather than on the unlabeled structure.

Watch

Extended reading notes

Core claim

The central claim is that context refactoring—recombining node attributes while holding interaction patterns fixed—creates augmented views that preserve node semantics and make anomalies more detectable, and that contrastive learning on these views lets a GNN exploit unlabeled data to learn discriminative embeddings even with very few labels. The paper also claims that encoding heterogeneous relation types separately and integrating them into message passing improves the model's ability to capture complex interaction semantics. Together these ingredients allow CRoC to outperform plain GNN baselines and existing GAD methods, with up to a 14% AUC gain, across seven real-world datasets.

Load-bearing premise

The load-bearing premise is that recomposing node attributes while keeping edges fixed preserves label-relevant meaning, so the contrastive views remain trustworthy; if refactoring mixes away the signal that marks an anomaly, the training learns from corrupted views and the reported gains would not transfer.

Editorial extensions

If this is right

CRoC should improve graph anomaly detection in settings where labeled anomalies are few and unlabeled nodes are abundant.
The method can be added to GNN architectures that handle heterogeneous edge types without needing extra supervision.
Because the augmentation preserves interaction patterns, the learned embeddings should be more robust to adversarial camouflage.
The contrastive objective lets the model use unlabeled data, so performance should degrade more gracefully as the label budget shrinks.

Reading between the lines

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

If context refactoring works, it suggests anomaly detection can be reframed as invariance to attribute mixing: anomalies are nodes whose features do not co-vary with their context the way normal nodes do; this is my inference, not a claim the paper states explicitly.
The same augmentation may transfer to other rare-class graph tasks, such as fraud detection in transaction networks, where labeled positives are scarce but unlabeled activity is abundant.
A stronger test would compare CRoC's augmentation against generic graph augmentations, like edge dropout or feature masking, under the same contrastive loss; if CRoC only wins on graphs with strong homophily, its advantage may be dataset-dependent rather than universal.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Here's my read. The one new thing in CRoC is the context-refactoring idea: build augmented graphs by taking attributes from one node and grafting them onto another's neighborhood while keeping edges fixed, then use contrastive learning with a few labeled anomalies. That is genuinely different from generic drop-and-perturb augmentations in GAD, and the separate encoding of heterogeneous relations is a sensible addition. If the experimental claims hold, this is a useful practical improvement for the limited-label regime, not a new theory, but that's fine.

The paper deserves a serious referee, but I want to flag what we can and cannot check. The full text I was sent is mojibake; the abstract is all I can read. So novelty and experimental detail rest on a single paragraph. That alone should prevent either a desk reject or a quick accept.

Soft spots, in proportion. The central invariance claim—recomposing attributes while preserving edges preserves node semantics—is asserted, not shown. For graph anomaly detection, many anomalies are exactly attribute-context mismatches; refactoring attributes can erase that signal or create synthetic mismatches that the contrastive head latches onto. A randomized permutation control (shuffle attributes across nodes while keeping edges) is needed to show the gain comes from the proposed refactoring rather than from any corruption-based augmentation. If the clean paper already has that ablation, my concern is answered; if not, that is the load-bearing question.

Second, the headline 'up to 14% AUC improvement' is a maximum over seven datasets, which can flatter the method. The abstract gives no per-dataset numbers, no error bars, no baseline list. Those may exist in the missing text, and I would not score the paper low for that alone, but I need to see them. The citation pattern cannot be checked from this version, though nothing in the abstract suggests a problem.

Bottom line: the idea is plausible and potentially useful; I cannot certify soundness from this version. Get a clean PDF, send it to review, and ask the referee specifically to check the permutation ablation and per-dataset results. This is not a paper to desk-reject on the abstract alone.

Referee Report

3 major / 3 minor

Summary. The paper proposes CRoC (Context Refactoring Contrast), a contrastive-learning framework for graph anomaly detection (GAD) under limited supervision. The method builds augmented graphs by recomposing node attributes while preserving edge structure, encodes heterogeneous relations separately, and trains a GNN with a joint supervised and contrastive objective. The abstract claims that CRoC achieves up to 14% AUC improvement over baseline GNNs and outperforms state-of-the-art GAD methods on seven real-world datasets. The full text supplied for review is heavily corrupted (mojibake), so none of the technical details, equations, experimental setup, or results tables could be read.

Significance. If the reported empirical gains are reproducible, the paper would offer a modest and practically useful improvement for GAD in low-label regimes, a setting where anomalies are rare and expensive to label. The conceptual idea of exploiting the inherent class imbalance of GAD to construct augmented contexts is interesting and distinguishes CRoC from generic graph augmentation methods. However, because the full text is unreadable, I cannot verify the soundness of the method, the correctness of the experimental comparisons, or the validity of the statistical conclusions. The paper provides no machine-checked proofs, released code, or reproducibility artifacts in the readable parts. The headline claim of 'up to 14% AUC improvement' is a maximum over datasets and is not accompanied by error bars or significance tests, so its robustness is unknown. The central invariance assumption behind the context-refactoring operation is stated without proof or ablation and may be particularly risky in GAD, where attribute-context mismatch is often the very definition of an anomaly.

major comments (3)

[Full text (after abstract)] The body of the manuscript is corrupted: the text after the abstract is mojibake, and no equation, algorithm, table, or experimental protocol is readable, so the technical derivation and the results in tables and figures cannot be verified. The header 'arXiv:2508.12281v3 [cs.CL] 9 Feb 2026' appears inside the manuscript, which is a different arXiv identifier from the submitted paper (2508.12278) and indicates a mismatched or concatenated PDF.
[Abstract] The abstract reports 'up to 14% AUC improvement' without error bars, number of runs, dataset statistics, or significance tests. Because this is a maximum over seven datasets, the headline could be driven by a single favorable comparison and does not by itself support the general claim that CRoC outperforms state-of-the-art methods under limited labels.
[Abstract (context refactoring description)] The central assumption that recomposing node attributes while preserving interaction patterns 'preserve[s] node semantics' is stated without proof or a random-permutation ablation. Since anomalies in GAD are often defined by attribute-context mismatch, refactoring could erase the anomaly signal or create synthetic mismatches that yield a shortcut; without a control experiment, the measured gain is not causally tied to the proposed context-refactoring mechanism.

minor comments (3)

[Whole manuscript] The manuscript lacks a reproducibility statement, and no reference to released code or data is visible in the abstract or any readable portion.
[Tables/figures (corrupted)] The dataset names, baseline configurations, and hyperparameters appear in the garbled tables but are indecipherable; a resubmission must include a clean version with these details.
[Abstract] The phrase 'simple yet effective' is not substantiated by any complexity analysis or runtime comparison in the readable parts of the manuscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity evidenced; abstract reports external benchmark evaluation and no fitted-input prediction chain.

full rationale

The available evidence (the abstract and the garbled full-text bytes) contains no step in which a claimed prediction is equivalent by construction to an input. CRoC is presented as a new training framework that refactors node contexts and combines limited labeled data with unlabeled data under a contrastive objective; the reported result is an AUC comparison on seven external GAD datasets. The contrastive objective is not a reformulation of the evaluation metric, and no parameter is fitted to the test labels and then renamed as a prediction. No load-bearing self-citation chain is visible, and no uniqueness theorem is invoked to force the method's choice. The full text is corrupted in the supplied material, so hidden benchmark tuning or post hoc selection cannot be checked, but the absence of such inspection is not itself circularity under the stated rules, which require quoting a specific reduction. Therefore the honest finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No free parameters or invented entities were identifiable from the abstract. The framework relies on three stated domain assumptions; a full audit is impossible with the corrupted full text.

assumptions (3)

domain assumption Anomalies are inherently rare, costly to label, and may actively camouflage their patterns.
Motivates the limited-label setting and the robustness target; taken as given from the problem domain rather than derived in the abstract.
ad hoc to paper Recomposing node attributes while preserving interaction patterns preserves node semantics.
Core design premise of CRoC's context refactoring; no proof or empirical validation is visible in the abstract.
domain assumption Contrastive learning on unlabeled data yields richer, more discriminative node embeddings for anomaly detection.
Standard self-supervised learning assumption; the paper relies on it to justify joint training, but the abstract provides no theoretical guarantee.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CRoC: Context Refactoring Contrast for Graph Anomaly Detection with Limited Supervision." pith.science (2026). https://pith.science/paper/VFTR4GI3

@misc{pith2026250812278,
  author       = {Pith},
  title        = {Pith review of: CRoC: Context Refactoring Contrast for Graph Anomaly Detection with Limited Supervision},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VFTR4GI3}},
  note         = {Machine review of arXiv:2508.12278}
}

read the original abstract

Graph Neural Networks (GNNs) are widely used as the engine for various graph-related tasks, with their effectiveness in analyzing graph-structured data. However, training robust GNNs often demands abundant labeled data, which is a critical bottleneck in real-world applications. This limitation severely impedes progress in Graph Anomaly Detection (GAD), where anomalies are inherently rare, costly to label, and may actively camouflage their patterns to evade detection. To address these problems, we propose Context Refactoring Contrast (CRoC), a simple yet effective framework that trains GNNs for GAD by jointly leveraging limited labeled and abundant unlabeled data. Different from previous works, CRoC exploits the class imbalance inherent in GAD to refactor the context of each node, which builds augmented graphs by recomposing the attributes of nodes while preserving their interaction patterns. Furthermore, CRoC encodes heterogeneous relations separately and integrates them into the message-passing process, enhancing the model's capacity to capture complex interaction semantics. These operations preserve node semantics while encouraging robustness to adversarial camouflage, enabling GNNs to uncover intricate anomalous cases. In the training stage, CRoC is further integrated with the contrastive learning paradigm. This allows GNNs to effectively harness unlabeled data during joint training, producing richer, more discriminative node embeddings. CRoC is evaluated on seven real-world GAD datasets with varying scales. Extensive experiments demonstrate that CRoC achieves up to 14% AUC improvement over baseline GNNs and outperforms state-of-the-art GAD methods under limited-label settings.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 32 canonical work pages

[1]

Bruna, W

J. Bruna, W. Zaremba, A. Szlam, and Y. LeCun. Spectral networks and locally connected networks on graphs. arXiv preprint arXiv:1312.6203, 2013

arXiv 2013
[2]

D. Chen, Y. Lin, W. Li, P. Li, J. Zhou, and X. Sun. Measuring and relieving the over-smoothing problem for graph neural networks from the topological view. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 3438--3445, 2020

work page 2020
[3]

J. Chen, G. Zhu, C. Yuan, and Y. Huang. Boosting graph anomaly detection with adaptive message passing. In The Twelfth International Conference on Learning Representations, 2024 a

work page 2024
[4]

N. Chen, Z. Liu, B. Hooi, B. He, R. Fathony, J. Hu, and J. Chen. Consistency training with learnable data augmentation for graph anomaly detection with limited supervision. In The Twelfth International Conference on Learning Representations, 2024 b

work page 2024
[5]

T. Chen, T. He, M. Benesty, V. Khotilovich, Y. Tang, H. Cho, K. Chen, R. Mitchell, I. Cano, T. Zhou, et al. Xgboost: extreme gradient boosting. R package version 0.4-2, 1 0 (4): 0 1--4, 2015

work page 2015
[6]

Y. Dou, Z. Liu, L. Sun, Y. Deng, H. Peng, and P. S. Yu. Enhancing graph neural network-based fraud detectors against camouflaged fraudsters. In Proceedings of the 29th ACM international conference on information & knowledge management, pages 315--324, 2020

work page 2020
[7]

W. Feng, J. Zhang, Y. Dong, Y. Han, H. Luan, Q. Xu, Q. Yang, E. Kharlamov, and J. Tang. Graph random neural networks for semi-supervised learning on graphs. Advances in neural information processing systems, 33: 0 22092--22103, 2020

work page 2020
[8]

Y. Gao, X. Wang, X. He, Z. Liu, H. Feng, and Y. Zhang. Alleviating structural distribution shift in graph anomaly detection. In Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, pages 357--365, 2023 a

work page 2023

Show all 40 references

[9]

Y. Gao, X. Wang, X. He, Z. Liu, H. Feng, and Y. Zhang. Addressing heterophily in graph anomaly detection: A perspective of graph spectrum. In Proceedings of the ACM Web Conference 2023, pages 1528--1538, 2023 b

2023
[10]

Hamilton, Z

W. Hamilton, Z. Ying, and J. Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017

2017
[11]

J. He, Q. Xu, Y. Jiang, Z. Wang, and Q. Huang. Ada-gad: Anomaly-denoised autoencoders for graph anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 8481--8489, 2024

2024
[12]

K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729--9738, 2020

2020
[13]

M. Jin, Y. Liu, Y. Zheng, L. Chi, Y.-F. Li, and S. Pan. Anemone: Graph anomaly detection with multi-scale contrastive learning. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 3122--3126, 2021

2021
[14]

T. N. Kipf and M. Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016

2016 arXiv
[15]

Q. Li, Y. He, C. Xu, F. Wu, J. Gao, and Z. Li. Dual-augment graph neural network for fraud detection. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, pages 4188--4192, 2022

2022
[16]

Y. Liu, X. Ao, Z. Qin, J. Chi, J. Feng, H. Yang, and Q. He. Pick and choose: a gnn-based imbalanced learning approach for fraud detection. In Proceedings of the web conference 2021, pages 3168--3177, 2021

2021
[17]

Z. Liu, Y. Dou, P. S. Yu, Y. Deng, and H. Peng. Alleviating the inconsistency problem of applying graph neural network to fraud detection. In Proceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval, pages 1569--1572, 2020

2020
[18]

X. Ma, J. Wu, S. Xue, J. Yang, C. Zhou, Q. Z. Sheng, H. Xiong, and L. Akoglu. A comprehensive survey on graph anomaly detection with deep learning. IEEE Transactions on Knowledge and Data Engineering, 35 0 (12): 0 12012--12038, 2021

2021
[19]

X. Ma, R. Li, F. Liu, K. Ding, J. Yang, and J. Wu. Graph anomaly detection with few labels: A data-centric approach. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2153--2164, 2024

2024
[20]

Qiao and G

H. Qiao and G. Pang. Truncated affinity maximization: One-class homophily modeling for graph anomaly detection. Advances in Neural Information Processing Systems, 36: 0 49490--49512, 2023

2023
[21]

A. Roy, J. Shu, J. Li, C. Yang, O. Elshocht, J. Smeets, and P. Li. Gad-nr: Graph anomaly detection via neighborhood reconstruction. In Proceedings of the 17th ACM international conference on web search and data mining, pages 576--585, 2024

2024
[22]

Schlichtkrull, T

M. Schlichtkrull, T. N. Kipf, P. Bloem, R. Van Den Berg, I. Titov, and M. Welling. Modeling relational data with graph convolutional networks. In The semantic web: 15th international conference, ESWC 2018, Heraklion, Crete, Greece, June 3--7, 2018, proceedings 15, pages 593--6...

2018
[23]

K. Sun, Z. Lin, and Z. Zhu. Multi-stage self-supervised learning for graph convolutional networks on graphs with few labeled nodes. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 5892--5899, 2020

2020
[24]

J. Tang, J. Li, Z. Gao, and J. Li. Rethinking graph neural networks for anomaly detection. In International Conference on Machine Learning, pages 21076--21089. PMLR, 2022

2022
[25]

J. Tang, F. Hua, Z. Gao, P. Zhao, and J. Li. Gadbench: Revisiting and benchmarking supervised graph anomaly detection. In Advances in Neural Information Processing Systems, volume 36, pages 29628--29653, 2023. URL https://proceedings.neurips.cc/paper_files/paper/2023/file/5eaa...

2023
[26]

Thakoor, C

S. Thakoor, C. Tallec, M. G. Azar, M. Azabou, E. L. Dyer, R. Munos, P. Veli c kovi \'c , and M. Valko. Large-scale representation learning on graphs via bootstrapping. arXiv preprint arXiv:2102.06514, 2021

2021 arXiv
[27]

Veli c kovi \'c , G

P. Veli c kovi \'c , G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y. Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017

2017 arXiv
[28]

Velickovic, W

P. Velickovic, W. Fedus, W. L. Hamilton, P. Li \`o , Y. Bengio, and R. D. Hjelm. Deep graph infomax. ICLR (Poster), 2 0 (3): 0 4, 2019

2019
[29]

Y. Wang, J. Zhang, S. Guo, H. Yin, C. Li, and H. Chen. Decoupling representation learning and classification for gnn-based anomaly detection. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval, pages 1239--1248, 2021

2021
[30]

B. Wu, X. Yao, B. Zhang, K.-M. Chao, and Y. Li. Splitgnn: Spectral graph neural network for fraud detection against heterophily. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 2737--2746, 2023

2023
[31]

Xiang, M

S. Xiang, M. Zhu, D. Cheng, E. Li, R. Zhao, Y. Ouyang, L. Chen, and Y. Zheng. Semi-supervised credit card fraud detection via attribute-driven graph representation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 14557--14565, 2023 a

2023
[32]

Xiang, M

S. Xiang, M. Zhu, D. Cheng, E. Li, R. Zhao, Y. Ouyang, L. Chen, and Y. Zheng. Semi-supervised credit card fraud detection via attribute-driven graph representation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 14557--14565, 2023 b

2023
[33]

S. Xie, W. C. Lau, et al. Cocos: Enhancing semi-supervised learning on graphs with unlabeled data via contrastive context sharing. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 4272--4280, 2022

2022
[34]

S. Xie, D. S. H. Tam, and W. C. Lau. Violin: virtual overbridge linking for enhancing semi-supervised learning on graphs with limited labels. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 4451--4459, 2023

2023
[35]

K. Xu, W. Hu, J. Leskovec, and S. Jegelka. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018

2018 arXiv
[36]

Z. Xu, L. Li, H. Li, Q. Sun, S. Hu, and R. Ji. Self-supervised graph representation learning for black market account detection. In Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, pages 330--338, 2023

2023
[37]

H. Yu, Z. Liu, and X. Luo. Barely supervised learning for graph-based fraud detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 16548--16557, 2024

2024
[38]

Zhang, J

G. Zhang, J. Wu, J. Yang, A. Beheshti, S. Xue, C. Zhou, and Q. Z. Sheng. Fraudre: Fraud detection dual-resistant to graph inconsistency and imbalance. In 2021 IEEE international conference on data mining (ICDM), pages 867--876. IEEE, 2021

2021
[39]

Zhang, M

H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017

2017 arXiv
[40]

Q. Zhou, Y. Chen, Z. Xu, Y. Wu, M. Pan, M. Das, H. Yang, and H. Tong. Graph anomaly detection with adaptive node mixup. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 3494--3504, 2024

2024

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

Bruna, W

J. Bruna, W. Zaremba, A. Szlam, and Y. LeCun. Spectral networks and locally connected networks on graphs. arXiv preprint arXiv:1312.6203, 2013

arXiv 2013

[2] [2]

D. Chen, Y. Lin, W. Li, P. Li, J. Zhou, and X. Sun. Measuring and relieving the over-smoothing problem for graph neural networks from the topological view. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 3438--3445, 2020

work page 2020

[3] [3]

J. Chen, G. Zhu, C. Yuan, and Y. Huang. Boosting graph anomaly detection with adaptive message passing. In The Twelfth International Conference on Learning Representations, 2024 a

work page 2024

[4] [4]

N. Chen, Z. Liu, B. Hooi, B. He, R. Fathony, J. Hu, and J. Chen. Consistency training with learnable data augmentation for graph anomaly detection with limited supervision. In The Twelfth International Conference on Learning Representations, 2024 b

work page 2024

[5] [5]

T. Chen, T. He, M. Benesty, V. Khotilovich, Y. Tang, H. Cho, K. Chen, R. Mitchell, I. Cano, T. Zhou, et al. Xgboost: extreme gradient boosting. R package version 0.4-2, 1 0 (4): 0 1--4, 2015

work page 2015

[6] [6]

Y. Dou, Z. Liu, L. Sun, Y. Deng, H. Peng, and P. S. Yu. Enhancing graph neural network-based fraud detectors against camouflaged fraudsters. In Proceedings of the 29th ACM international conference on information & knowledge management, pages 315--324, 2020

work page 2020

[7] [7]

W. Feng, J. Zhang, Y. Dong, Y. Han, H. Luan, Q. Xu, Q. Yang, E. Kharlamov, and J. Tang. Graph random neural networks for semi-supervised learning on graphs. Advances in neural information processing systems, 33: 0 22092--22103, 2020

work page 2020

[8] [8]

Y. Gao, X. Wang, X. He, Z. Liu, H. Feng, and Y. Zhang. Alleviating structural distribution shift in graph anomaly detection. In Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, pages 357--365, 2023 a

work page 2023