Pith. sign in

REVIEW 3 major objections 3 minor 41 references

GTHNA claims that combining a local-global Transformer encoder, a memory of normal patterns, and multi-scale reconstruction yields state-of-the-art unsupervised node anomaly detection.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 17:28 UTC pith:S6PLUJKK

load-bearing objection Solid method-level combination with a self-referential memory bootstrap and a thinner empirical edge than the abstract claims; worth a serious referee, not a desk reject. the 3 major comments →

arxiv 2509.10869 v1 pith:S6PLUJKK submitted 2025-09-13 cs.LG cs.AI

GTHNA: Local-global Graph Transformer with Memory Reconstruction for Holistic Node Anomaly Evaluation

Mingkang Li , Xuexiong Luo , Yue Zhang , Yaoyang Li , Fu Lin This is my paper
classification cs.LG cs.AI
keywords unsupervised graph anomaly detectiongraph transformermemory modulemulti-scale reconstructionnode anomaly evaluationreconstruction errorpositional encodingover-smoothing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes GTHNA, an unsupervised framework for detecting anomalous nodes in attributed graphs. It argues that existing graph-convolutional detectors suffer from over-smoothing, which makes normal and anomalous representations indistinguishable, and that reconstruction-based methods let anomalies contaminate the reconstruction process. GTHNA addresses both by encoding nodes with a Transformer that fuses local k-hop subgraph summaries with Laplacian positional encodings, by reconstructing nodes through a memory module that stores prototypical normal patterns, and by scoring anomalies with a combination of multi-scale reconstruction errors and memory-matching distances. The paper reports that GTHNA outperforms nine baselines on seven benchmarks, achieving the highest AUC on six of them. The central claim is that holistic, memory-guided, multi-scale evaluation substantially improves unsupervised graph anomaly detection.

Core claim

The central claim is that node anomaly detection in attributed graphs can be substantially improved by evaluating each node holistically: encode with a local-global Transformer whose input combines a GCN-computed k-hop subgraph representation and a Laplacian-based positional encoding; then reconstruct the node through a memory module of trainable prototypes that are updated using the 80% of nodes with smallest reconstruction error; and finally score each node as the weighted sum of self-reconstruction error (attributes and edges), neighborhood-reconstruction error (degree and neighbor-feature distribution), and distance to the nearest memory item. The paper argues that the memory module supp

What carries the argument

The memory-guided decoder is the load-bearing component. It stores a set of prototypical patterns of normal nodes; each node embedding reads the memory by a softmax-weighted average of memory items, and the memory items are updated by aggregating embeddings from the 80% of nodes with the lowest reconstruction error. The paper claims this both makes reconstruction of normal nodes more accurate and amplifies reconstruction error for anomalies. The other two components are the local-global Transformer encoder, which fuses a GCN-computed k-hop subgraph representation with Laplacian positional encoding, and a multi-scale reconstruction loss that reconstructs node attributes, edges, and neighborho

Load-bearing premise

The framework assumes that the 80% of nodes with the smallest reconstruction error are normal, but that error is computed through the memory module itself, so the memory can only learn what it already considers normal, and there is no independent signal that the pseudo-normal set is actually anomaly-free.

What would settle it

Train GTHNA on a graph where anomalies are deliberately constructed to have low reconstruction error under the model's own scoring (for example, by copying normal feature distributions but altering only a non-reconstructed attribute), and check whether the memory items drift toward the anomalies. If the nearest prototypes for anomalous nodes become highly similar, or if AUC drops below the no-memory ablation, the memory-guided improvement claim is falsified. Alternatively, replace the U_h selection with a random 80% of nodes; if the performance gap persists, the memory's normality assumption i

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If GTHNA's central claim holds, graph anomaly detectors can be built that remain effective on graphs with organic, non-injected anomalies, not just synthetic ones.
  • The memory-module design implies that explicit modeling of normality, rather than only reconstruction error, is a useful direction for graph anomaly detection.
  • The multi-scale reconstruction (node self plus neighborhood) implies that neighborhood statistics carry signal beyond edge reconstruction.
  • The local-global Transformer encoder implies that over-smoothing can be mitigated without sacrificing local structural information.
  • The reported six-of-seven best AUC suggests the method generalizes across domains as different as citation networks, social networks, and review graphs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension the authors do not run: replace the self-referential normal-node selection (low reconstruction error) with an independent normality prior, such as degree or local-density heuristics, and compare memory quality. If performance drops when the selection is independent, the reported gains may come more from the Transformer encoder than from the memory.
  • Another implication left implicit: because the memory is updated from the 80% lowest-error nodes, the method's success depends on anomalies being a small minority and on their reconstruction errors being initially higher than those of normal nodes. On datasets where anomalies are designed to look normal, the memory could absorb anomalous patterns and the framework would collapse to a standard auto
  • The paper's complexity analysis includes an O(n^3) cost for Laplacian eigendecomposition, which it notes can be precomputed; for very large graphs, approximate positional encodings would be needed, as the paper's own future-work section flags.
  • One could probe the memory directly: after training, retrieve the nearest memory item for each anomalous node. If anomalies have near-zero distance to some memory item, then the memory has memorized anomalies rather than normality, undercutting the claimed mechanism.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. GTHNA proposes an unsupervised graph anomaly detection framework combining a local-global graph Transformer encoder (k-hop subgraph GCN plus Laplacian absolute positional encoding), a memory module intended to store prototypical normal-node patterns, and a multi-scale reconstruction objective covering node attributes, edges, and neighborhood distributions. The final anomaly score is a weighted combination of reconstruction errors and memory-matching distance. Experiments on seven attributed graph benchmarks report AUC; GTHNA obtains the best mean AUC on six of seven datasets, and ablations attribute performance gains to the structure extractor and the memory module. The manuscript includes a code link, ten-seed results, and a complexity analysis.

Significance. If the causal claims were established, the paper would make a useful contribution: it offers a concrete architecture for combining local and global structural encoding with memory-guided reconstruction for graph anomaly detection, and it reports results across a broad set of benchmarks. The manuscript's strengths include code release, multiple random seeds, ablation studies, and an explicit complexity analysis. However, the central mechanism is not yet supported. The memory update rule uses reconstruction errors that themselves flow through the memory, so the 'normal pattern' signal is self-referential; the paper provides no evidence that the pseudo-normal set remains uncontaminated. In addition, the headline empirical claim is weakened by a loss on Weibo and by improvements within one standard deviation on Reddit and Citeseer. The result may be correct, but the current evidence does not establish the claimed memory-guided suppression of anomaly interference.

major comments (3)
  1. [Table 3] The memory module is the core contribution, but its training signal is circular. U_h is defined as nodes with smaller reconstruction errors (Eq. 13), while those reconstruction errors are computed through the memory-guided decoder: h_hat_i is the weighted memory read in Eq. (11), the concatenation of h_a_i and h_hat_i is fed to the decoder, and Eqs. (18)-(19) produce the reconstruction losses that determine U_h. Thus a node can enter U_h precisely because it matches a currently stored memory item, and then reinforce that item for the next update. Figure 5 reports only the post-training similarity of final memory items to normal vs anomalous nodes; it does not rule out contamination during training. This is load-bearing because the ablation in §5.3 credits the memory module with a large performance gain. Requested: a label-based audit on the injected-anomaly datasets, e.g., the fraction o
  2. [§4.4, Eqs. (22)-(23)] The claim that GTHNA outperforms state-of-the-art methods is not fully supported. On Weibo, DiffGAD achieves 93.40±0.30 vs GTHNA 92.38±0.15, so GTHNA does not have the best mean AUC there. On Reddit, GTHNA 59.21±1.63 vs ANEMONE 57.99±0.99 and GAD-NR 57.99±1.67. On Citeseer, GTHNA 93.62±0.12 vs SL-GAD 93.24±0.92. In the latter two cases the standard-deviation intervals overlap, and the reported means alone are not robust evidence of superiority. Please provide per-seed paired comparisons or significance tests, such as paired t-test or Wilcoxon, for the headline comparisons, and adjust the abstract and introduction wording accordingly.
  3. [Minor] The final anomaly score in Eq. (23) is, up to the additive memory-separateness term, the same loss minimized in Eq. (22). The paper does not report a held-out split; training and evaluation appear to use the same nodes. If the network can reduce the training loss without learning meaningful structure, the reported AUC can partly reflect optimization of the evaluation criterion. Please clarify the evaluation protocol, including transductive vs inductive, checkpoint selection, and early stopping, and if feasible include an evaluation on held-out nodes or justify the transductive reconstruction setting with additional overfitting analysis.
minor comments (3)
  1. [Eq. (16)] The bracketed term writes lambda_n L^a_i, but the simplified form uses lambda_n L_n; the first expression should be lambda_n L^n_i. Also, D(a_i - a_hat_i) and D(x_i - x_hat_i) should presumably be written as distance functions of two arguments, e.g., D(a_i, a_hat_i).
  2. [§5.1] The sum over neighbors is written as sum over v_i in H_i; it should be over v_j in H_i. The formula for omega_i also divides by c_i - 1, which is undefined for isolated nodes with c_i = 1. Please handle zero- or one-neighbor cases explicitly.
  3. [Abstract] The baseline comparison gives dataset statistics and result tables but no configuration details for the baselines, such as hyperparameters, number of epochs, or early stopping. Please report these, or state that the baselines' official or standard settings were used, to make the comparison reproducible.

Circularity Check

1 steps flagged

Memory-guided 'normal pattern' bootstrap is self-referential, but the benchmarked AUC claim remains independent; partial circularity.

specific steps
  1. self definitional [Section 4.2 (Memory-guided Decoder), Eqs. 10–13 and Eqs. 18–19]
    "During unsupervised training, the memory module is dynamically updated by identifying nodes with relatively low reconstruction errors and treating them as pseudo-normal samples. ... We use the nodes with smaller reconstruction errors, denoted as U_h, and assume that U_h is the set of normal nodes, use it to update the memory items."

    The 'normal' set U_h is selected using reconstruction errors L^s_i, L^a_i, L^n_i (Eqs. 18–19), which are computed from the decoder input \bar{h}_i = [h^a_i; \hat{h}_i], where \hat{h}_i = Σ_j S^Q_{i,j} M_j (Eq. 11) is a weighted read of the memory items. The same memory items are then updated by Eq. 13 using exactly the nodes in U_h. Thus normality is defined by reconstruction errors that are themselves functions of the memory, while the memory is defined by nodes deemed normal under those errors. There is no independent normality signal; the claimed 'prototypical patterns of normal nodes' are a self-referential fixed point. The final anomaly score (Eq. 23) is the same reconstruction/memory objective (Eq. 22), so the scoring function is the training loss by construction.

full rationale

GTHNA's headline result is an empirical AUC comparison against nine baselines on seven benchmark datasets. These AUC numbers are computed with held-out ground-truth anomaly labels, so the central performance claim does not reduce to the model's own training objective; it is independent evidence. The one substantive circularity concern is inside the memory module: the set U_h of 'normal' nodes is selected by low reconstruction error, but reconstruction error is computed through the memory-read decoder (Eq. 11) that the same U_h updates (Eq. 13). This makes the memory's notion of normality self-referential, and the paper's RQ3 analysis (Fig. 5) only checks the final state, not whether the loop remained uncontaminated during training. Because this affects a component of the method (memory-guided reconstruction) rather than the benchmarked AUC claim itself, and because the Transformer/multi-scale/neighborhood-reconstruction components are independently defined, the overall circularity is partial: score 4.

Axiom & Free-Parameter Ledger

9 free parameters · 7 axioms · 1 invented entities

The central claim depends on a standard problem setting (undirected attributed graphs), on the rarity and pattern-ability of anomalies, and on a self-referential memory bootstrap. The method introduces no new physical entities, but the memory module is a learned construct whose 'normality' is only defined through the reconstruction errors it helps produce. Hyperparameters lambda, k, m, beta, U_h, and d_h are set manually with no described validation split.

free parameters (9)
  • lambda_s = 1.0
    Loss weight for node self-reconstruction (structure + attribute), set manually in Implementation Details.
  • lambda_n = 0.01
    Loss weight for neighborhood reconstruction, set manually.
  • lambda_m = 0.01
    Loss weight for memory matching and separation; Fig 7 shows AUC varies with this value on Citeseer, and no held-out tuning is described.
  • k = 2
    Hop count for the local subgraph extractor, chosen manually.
  • U_h proportion = 80%
    Proportion of lowest reconstruction-error nodes treated as normal for memory updates; assumes anomaly rate is below 20%.
  • m = 512
    Number of memory items, set manually.
  • beta = 0.5
    Margin in the memory separateness loss, set manually.
  • d_h = 128
    Node embedding dimension, set manually.
  • attention heads = 4
    Number of attention heads in the Transformer encoder, set manually.
axioms (7)
  • domain assumption The graph is undirected and attributed (Definition 3.1).
    The method and evaluation only cover undirected attributed graphs; directed or dynamic graphs are outside scope.
  • domain assumption Most nodes are normal, so the 80% lowest reconstruction-error nodes can be treated as normal (Section 4.2).
    If anomalies exceed 20% or are systematically low-error, the memory update is poisoned.
  • domain assumption Normal nodes can be represented by m prototypical memory items, and deviations from these prototypes indicate anomaly (Section 4.2).
    No evidence is given that real normal-node patterns cluster into 512 prototypes.
  • domain assumption Reconstruction error is a valid anomaly score (Section 4.4, following DOMINANT [3]).
    Adopted from prior work; if anomalies are also reconstructible, the score fails.
  • standard math Eigendecomposition of the normalized Laplacian yields a useful absolute positional encoding (Eq 3).
    Standard spectral positional encoding, not questioned.
  • standard math The KL divergence and L2 distances in Eqs (18)-(19) are appropriate for comparing neighborhood distributions.
    Standard distance choices, no derivation given.
  • ad hoc to paper Unsupervised training minimizes the same loss later used as anomaly score (Eq 22 vs Eq 23).
    The model is trained to minimize anomaly scores of all nodes, including anomalies; the score is not independent of the training objective.
invented entities (1)
  • Memory items M_j (j=1..m) no independent evidence
    purpose: Store prototypical normal-node patterns and guide reconstruction; read by cosine attention and updated by low-error nodes.
    Learned parameter vectors adapted from [22]; no external handle outside this paper, and their 'normality' is inferred from the self-referential reconstruction error.

pith-pipeline@v1.3.0-alltime-deepseek · 22455 in / 15855 out tokens · 162339 ms · 2026-08-04T17:28:58.978747+00:00 · methodology

0 comments
read the original abstract

Anomaly detection in graph-structured data is an inherently challenging problem, as it requires the identification of rare nodes that deviate from the majority in both their structural and behavioral characteristics. Existing methods, such as those based on graph convolutional networks (GCNs), often suffer from over-smoothing, which causes the learned node representations to become indistinguishable. Furthermore, graph reconstruction-based approaches are vulnerable to anomalous node interference during the reconstruction process, leading to inaccurate anomaly detection. In this work, we propose a novel and holistic anomaly evaluation framework that integrates three key components: a local-global Transformer encoder, a memory-guided reconstruction mechanism, and a multi-scale representation matching strategy. These components work synergistically to enhance the model's ability to capture both local and global structural dependencies, suppress the influence of anomalous nodes, and assess anomalies from multiple levels of granularity. Anomaly scores are computed by combining reconstruction errors and memory matching signals, resulting in a more robust evaluation. Extensive experiments on seven benchmark datasets demonstrate that our method outperforms existing state-of-the-art approaches, offering a comprehensive and generalizable solution for anomaly detection across various graph domains.

Figures

Figures reproduced from arXiv: 2509.10869 by Fu Lin, Mingkang Li, Xuexiong Luo, Yaoyang Li, Yue Zhang.

Figure 1
Figure 1. Figure 1: Normal node 𝑢 and anomalous node 𝑣 have distinct topological structures. However, when encoded by a graph Transformer using absolute positional encoding (shortest path), they get same positional encoding. This erases struc￾tural differences between nodes, weakening the model’s abil￾ity to detect structure anomalies at the node level. with absolute positional encoding may fail to capture the structural diff… view at source ↗
Figure 3
Figure 3. Figure 3: The overview of GTHNA. The input nodes will be encoded by the local-global graph Transformer to get discriminate [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of ablation experiment results. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Cosine similarity comparison between memory [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Impacts of loss weights 𝜆𝑠 , 𝜆𝑛 and 𝜆𝑚 on model performance. 6 Conclusion and Outlook To address the over-smoothing issue in node representation learn￾ing with GCNs and the interference of anomalous nodes in graph reconstruction, we propose a holistic memory-Transformer frame￾work for anomaly detection (GTHNA). By enhancing the node posi￾tional encoding, our method improves node representation distinc￾tive… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 5 linked inside Pith

  1. [1]

    Deli Chen, Yankai Lin, Wei Li, Peng Li, Jie Zhou, and Xu Sun. 2020. Measuring and relieving the over-smoothing problem for graph neural networks from the topological view. InAAAI, Vol. 34. 3438–3445

    2020

  2. [2]

    Dexiong Chen, Leslie O’Bray, and Karsten Borgwardt. 2022. Structure-aware transformer for graph representation learning. InICML. PMLR, 3469–3489

    2022

  3. [3]

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

    2019

  4. [4]

    Xusheng Du, Jiong Yu, Zheng Chu, Lina Jin, and Jiaying Chen. 2022. Graph autoencoder-based unsupervised outlier detection.Information Sciences608 (2022), 532–550

    2022

  5. [5]

    Jingcan Duan, Siwei Wang, Pei Zhang, En Zhu, Jingtao Hu, Hu Jin, Yue Liu, and Zhibin Dong. 2023. Graph anomaly detection via multi-scale contrastive learning networks with augmented view. InAAAI, Vol. 37. 7459–7467

    2023

  6. [6]

    Ganghui Fan, Yong Ma, Xiaoguang Mei, Fan Fan, Jun Huang, and Jiayi Ma

  7. [7]

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

    2020

  8. [8]

    Zehao Guo, Nannan Wu, Yiming Zhao, and Wenjun Wang. 2024. TransGAD: A Transformer-Based Autoencoder for Graph Anomaly Detection. InDASFAA. Springer, 269–284

    2024

  9. [9]

    Junwei He, Qianqian Xu, Yangbangyan Jiang, Zitai Wang, and Qingming Huang

  10. [10]

    Yihong Huang, Liping Wang, Fan Zhang, and Xuemin Lin. 2023. Unsupervised graph outlier detection: Problem revisit, new insight, and superior method. In ICDE. IEEE, 2565–2578

    2023

  11. [11]

    Md Shamim Hussain, Mohammed J Zaki, and Dharmashankar Subramanian

  12. [12]

    Ming Jin, Yixin Liu, Yu Zheng, Lianhua Chi, Yuan-Fang Li, and Shirui Pan. 2021. Anemone: Graph anomaly detection with multi-scale contrastive learning. In CIKM. 3122–3126

    2021

  13. [13]

    Parisa Kaghazgaran, James Caverlee, and Anna Squicciarini. 2018. Combating crowdsourced review manipulators: A neighborhood-based approach. InWSDM. 306–314

    2018

  14. [14]

    Edge-augmented graph transformers: Global self-attention is enough for graphs.arXiv preprint arXiv:2108.033483 (2021)

    Pith/arXiv arXiv 2021

  15. [15]

    Srijan Kumar, Xikun Zhang, and Jure Leskovec. 2019. Predicting dynamic em- bedding trajectory in temporal interaction networks. InSIGKDD. 1269–1278

    2019

  16. [16]

    Eren Kurshan, Hongda Shen, and Haojie Yu. 2020. Financial crime & fraud detection using graph computing: Application considerations & outlook. In TransAI. IEEE, 125–130

    2020

  17. [17]

    Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907(2016)

    Pith/arXiv arXiv 2016

  18. [18]

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

  19. [19]

    Qingcheng Lu, Nannan Wu, Yiming Zhao, Wenjun Wang, and Quannan Zu. 2024. Enhancing Multi-view Contrastive Learning for Graph Anomaly Detection. In DASFAA. Springer, 236–251

    2024

  20. [20]

    Jinghan Li, Yuan Gao, Jinda Lu, Junfeng Fang, Congcong Wen, Hui Lin, and Xiang Wang. 2024. Diffgad: A diffusion-based unsupervised graph anomaly detector. arXiv preprint arXiv:2410.06549(2024)

    Pith/arXiv arXiv 2024

  21. [21]

    Xiaoxiao Ma, Jia Wu, Shan Xue, Jian Yang, Chuan Zhou, Quan Z Sheng, Hui Xiong, and Leman Akoglu. 2021. A comprehensive survey on graph anomaly detection with deep learning.IEEE Transactions on Knowledge and Data Engineering35, 12 (2021), 12012–12038

    2021

  22. [22]

    Anomaly detection on attributed networks via contrastive self-supervised learning.IEEE Transactions on Neural Networks and Learning Systems33, 6 (2021), 2378–2392

    2021

  23. [23]

    Sanjeev Rao, Anil Kumar Verma, and Tarunpreet Bhatia. 2021. A review on social spam detection: Challenges, open issues, and future directions.Expert Systems with Applications186 (2021), 115742

    2021

  24. [24]

    Xuexiong Luo, Jia Wu, Amin Beheshti, Jian Yang, Xiankun Zhang, Yuan Wang, and Shan Xue. 2022. Comga: Community-aware attributed graph anomaly detec- tion. InWSDM. 657–665

    2022

  25. [25]

    Amit Roy, Juan Shu, Jia Li, Carl Yang, Olivier Elshocht, Jeroen Smeets, and Pan Li. 2024. Gad-nr: Graph anomaly detection via neighborhood reconstruction. In WSDM. 576–585

    2024

  26. [26]

    Hyunjong Park, Jongyoun Noh, and Bumsub Ham. 2020. Learning memory- guided normality for anomaly detection. InCVPR. 14372–14381

    2020

  27. [27]

    Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. 2018. Self-attention with relative position representations.arXiv preprint arXiv:1803.02155(2018)

    Pith/arXiv arXiv 2018

  28. [28]

    Shebuti Rayana and Leman Akoglu. 2015. Collective opinion spam detection: Bridging review networks and metadata. InSIGKDD. 985–994

    2015

  29. [29]

    Jie Tang, Jing Zhang, Limin Yao, Juanzi Li, Li Zhang, and Zhong Su. 2008. Arnet- miner: extraction and mining of academic social networks. InSIGKDD. 990–998

    2008

  30. [30]

    Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi-Rad. 2008. Collective classification in network data.AI Magazine29, 3 (2008), 93–93

    2008

  31. [31]

    Petar Veličković. 2023. Everything is connected: Graph neural networks.Current Opinion in Structural Biology79 (2023), 102538

    2023

  32. [32]

    Jianheng Tang, Jiajin Li, Ziqi Gao, and Jia Li. 2022. Rethinking graph neural networks for anomaly detection. InICML. PMLR, 21076–21089

    2022

  33. [33]

    Yirui Wu, Dabao Wei, and Jun Feng. 2020. Network attacks detection meth- ods based on deep learning techniques: a survey.Security and Communication Networks2020, 1 (2020), 8872923

    2020

  34. [34]

    Lei Tang and Huan Liu. 2009. Relational learning via latent social dimensions. In SIGKDD. 817–826

    2009

  35. [35]

    Tong Zhao, Chuchen Deng, Kaifeng Yu, Tianwen Jiang, Daheng Wang, and Meng Jiang. 2020. Error-bounded graph anomaly loss for gnns. InCIKM. 1873–1882

    2020

  36. [36]

    Xinyi Wu, Amir Ajorlou, Zihui Wu, and Ali Jadbabaie. 2024. Demystifying oversmoothing in attention-based graph neural networks.Advances in Neural Information Processing Systems36 (2024)

    2024

  37. [37]

    Qinghai Zhou, Yuzhong Chen, Zhe Xu, Yuhang Wu, Menghai Pan, Mahashweta Das, Hao Yang, and Hanghang Tong. 2024. Graph Anomaly Detection with Adaptive Node Mixup. InCIKM. 3494–3504

    2024

  38. [38]

    Jiawei Zhang, Haopeng Zhang, Congying Xia, and Li Sun. 2020. Graph-bert: Only attention is needed for learning graph representations.arXiv preprint arXiv:2001.05140(2020)

    Pith/arXiv arXiv 2020

  39. [40]

    Yu Zheng, Ming Jin, Yixin Liu, Lianhua Chi, Khoa T Phan, and Yi-Ping Phoebe Chen. 2021. Generative and contrastive self-supervised learning for graph anom- aly detection.IEEE Transactions on Knowledge and Data Engineering35, 12 (2021), 12220–12233

    2021

  40. [2021]

    Hyperspectral anomaly detection with robust graph autoencoders.IEEE Transactions on Geoscience and Remote Sensing60 (2021), 1–14

    2021

  41. [2024]

    InAAAI, Vol

    Ada-gad: Anomaly-denoised autoencoders for graph anomaly detection. InAAAI, Vol. 38. 8481–8489