Pith. sign in

REVIEW 3 major objections 5 minor 50 references

DP-DGAD: A Generalist Dynamic Graph Anomaly Detector with Dynamic Prototypes

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A single dynamic-graph anomaly detector can generalize across eight unlabeled target domains by maintaining a memory buffer of 'dynamic prototypes' that track how normal and anomalous patterns evolve over time, and the paper reports…

desk verdict New problem framing but the target adaptation protocol leaks test anomalies into the model, so the reported gains are not interpretable. read the letter →

arxiv 2508.00664 v1 pith:LUP3PLPO submitted 2025-08-01 cs.LG

Jialun Zheng , Jie Liu , Jiannong Cao , Xiao Wang , Hanchen Yang , Yankai Chen , Philip S. Yu This is my paper · ORCID
classification cs.LG
keywords dynamicgraphanomalydetectiongeneralistprototypesmemorybufferpseudo-labelingdomainshiftneuralnetworksself-supervisedadaptation
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 dynamic graph anomaly detection can be made generalist: a single model pretrained on labeled source graphs can detect anomalous edges in new, unlabeled graphs from other domains. The proposed mechanism represents 'normal' and 'anomalous' as dynamic prototypes — evolving embeddings extracted from temporal ego-graphs — kept in a memory buffer that retains domain-agnostic patterns while admitting domain-specific ones. A distribution-based scorer compares incoming edges to the prototype statistics, and low-entropy detections on target graphs become pseudo-labels for self-supervised adaptation. On eight target datasets spanning finance, communication, and citation networks, the paper reports consistently higher AUROC and AUPRC than all baselines, with the largest gains at low anomaly ratios.

What carries the argument

The load-bearing object is the dynamic prototype pair ($p_n$, $p_a$) stored in a memory buffer whose size is set to 10% of the training data. Each pair is aligned to normal and anomalous edge embeddings via an alignment loss; a difference score (mean Euclidean distance between the two prototypes) keeps only the most distinguishable pairs, and a ranking score $s_r = \lambda_d s_d - \lambda_e s_e$ replaces the least general pairs when a new source domain is encountered. The same prototypes feed a distribution-based scorer that tracks mean and covariance, producing scores that double as confidence estimates for entropy-based pseudo-labeling during target adaptation.

What would settle it

Run the target-domain protocol with a temporal split: fine-tune on early edges of each target graph, inject anomalies only into later edges, and test only on those later edges; if AUROC collapses to novelty-detection levels, the reported gains reflect edge novelty rather than learned anomaly semantics.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that a prototype-based memory system can carry anomaly semantics across domains in evolving graphs. DP-DGAD learns two sets of prototypes (normal and abnormal) from labeled source graphs, updates them with edge-level temporal information, and stores the most discriminative pairs in a bounded buffer. When a new domain arrives, prototypes already in the buffer that are both internally discriminative and similar to the new domain's representations are retained as domain-agnostic, while lower-ranked pairs are replaced with new domain-specific prototypes. Anomaly scores come from the gap between an edge's similarity to the abnormal distribution and its similarity to the normal distribution, using rolling mean and covariance statistics. The paper reports that this combination — dynamic prototypes plus confidence-gated pseudo-labeling — achieves the best AUROC and AUPRC on all eight target datasets under three anomaly injection ratios.

Load-bearing premise

The evaluation protocol fine-tunes on each target dataset using only normal edges and then tests on the same dataset after injecting anomalous edges, without separating the adaptation edges from the test edges, so the reported scores may reflect the model detecting simply that the injected edges are new rather than that it learned what makes an edge anomalous.

Editorial extensions

If this is right

  • A generalist dynamic graph anomaly detector is feasible without any target labels, reducing the cost of deploying detectors in new domains.
  • Low-anomaly-ratio settings — where anomalous edges are rare — benefit most, which is the regime typical of fraud and intrusion detection.
  • The memory-buffer prototype mechanism is a drop-in scoring head that could be attached to future backbone encoders for evolving graphs.
  • Confidence-based pseudo-labeling, rather than distance- or similarity-based selection, is what makes target adaptation work, suggesting entropy is a reliable reliability signal.
  • If the reported gains hold, the field's 'one model per domain' paradigm for dynamic graphs gives way to a single pretrained generalist that can be adapted on the fly.

Reading between the lines

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

  • The evaluation protocol may let novelty detection masquerade as anomaly detection: injected anomalous edges connect previously disconnected nodes, so the model could be flagging 'unseen anywhere before' rather than 'anomalous by learned semantics'; a temporal holdout split would separate these.
  • The buffer behaves like a continual-learning memory; a natural extension would test whether sequential exposure to many target domains causes catastrophic forgetting of the earliest source patterns.
  • The distribution-based scorer with rolling mean and covariance resembles a Mahalanobis-style metric; coupling prototype updates to online covariance estimation could be tested on streaming settings where anomaly semantics drift within a single domain.
  • One could extend the pseudo-labeling stage to a teacher-student or uncertainty-calibrated variant to see whether gains hold when target domains are far out-of-distribution.
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.

Referee Report

3 major / 5 minor

Summary. The paper proposes DP-DGAD, a generalist dynamic graph anomaly detector that pretrains on two labeled source graphs (Wikipedia and MOOC) and adapts to eight unlabeled target graphs. The method extracts temporal ego-graph representations, maintains dynamic normal/abnormal prototypes in a memory buffer, scores edges by their similarity to prototype distributions, and performs target-domain adaptation through confidence-based pseudo-labeling. The central empirical claim, based on Table 2, is that DP-DGAD consistently achieves the highest AUROC and AUPRC across all eight target datasets and three anomaly ratios.

Significance. If the empirical results were valid, this would be a meaningful contribution: cross-domain transfer for dynamic graph anomaly detection is an important and under-studied problem, and the dynamic-prototype memory buffer plus entropy-based pseudo-labeling are reasonable design choices. The paper also provides an anonymous code link and fairly extensive ablations. However, the evaluation protocol as described undermines the main claim, because target-domain adaptation and testing appear to overlap on the same edges, so the reported gains in Table 2 are not interpretable as evidence of learned, transferable anomaly semantics.

major comments (3)
  1. [Section 5.1 vs Section 4.3 and Algorithm 2] There is a direct contradiction about what the target datasets contain during adaptation. Section 5.1 states that the eight target datasets 'contain only normal edges and no abnormal ones' during fine-tuning and that testing is performed on 'the same eight target datasets, but with injected anomalies.' Section 4.3 and Algorithm 2, however, describe selecting the top N_con 'abnormal and normal detection pairs' on target data as pseudo-labels. If the target data used for pseudo-labeling are normal-only, abnormal pseudo-labels are impossible; if the target data include the injected anomalies, then the model is adapted with knowledge of the test anomalies. Either way, the paper's label-free transfer claim is unsupported as written. The authors must state explicitly whether pseudo-labeling runs on the normal-only target data or on the data with injected anomalies, and if a temporal or edge-disjoint split separates adaptation from testing, the text and Algorithm 2 must describe it.
  2. [Appendix A.1 and Table 2] The anomaly injection protocol also creates a novelty-detection confound independent of the pseudo-labeling issue. Appendix A.1 says anomalies are injected by linking 'original disconnected nodes.' Because the model is fine-tuned on the normal edges of the same target graph, every injected anomalous edge connects a pair of nodes that were disconnected during training, making those edges the only genuinely unseen edges in the test set. High AUROC could therefore reflect edge novelty rather than anomaly semantics learned from the source graphs. The 10-20 point gains reported in Table 2 are not interpretable as evidence for the generalist-DGAD claim unless the evaluation uses a held-out temporal split or a target graph not seen during adaptation.
  3. [Section 5.1, Implementation Details] The description of the evaluation protocol needs to be precise about what data are used for each stage. In particular, the sentence 'they are fine-tuned on eight target datasets, which contain only normal edges and no abnormal ones. Following this, testing is performed on the same eight target datasets, but with injected anomalies' is too ambiguous to support the central claim. The paper should specify the exact split between adaptation edges and test edges, the timestamps used for each, and how pseudo-labels are generated without using test labels. Without this, the reported state-of-the-art results cannot be reproduced or trusted.
minor comments (5)
  1. [Equations (13)-(14)] The covariance update for the normal distribution in Eq. (13) omits the momentum factor on the previous covariance term, while Eq. (14) includes it: Σ_{n,t} = Σ_{n,t-1} + (1-α) ... versus Σ_{a,t} = αΣ_{a,t-1} + (1-α) .... This is likely a typo, but it should be fixed for reproducibility.
  2. [Algorithms 1 and 2] The equation references inside Algorithm 1 and Algorithm 2 are inconsistent with the numbering in the text; for example, Algorithm 1 refers to 'Eq.17 to Eq.18' for computing anomaly scores, but Eqs. 15-16 define the similarity scores and Eq. 17 defines the BCE loss. The referenced equation numbers should be corrected.
  3. [Section 5.5] In the hyperparameter analysis, the text says 'covering broader aspects of the distribution of animal patterns'; this should be 'anomalous patterns.'
  4. [Throughout] There are several typographical errors, including 'DP-GDAD' in the introduction, 'psuedo' in Section 5.4, and 'University of Illinons' in the author affiliation. A careful proofread is needed.
  5. [Section 5.1] The memory buffer size M and the number of confident detections N_con are both set to '10% of the data size,' but it is not specified whether this refers to the source dataset size, the target dataset size, or something else; this should be clarified.

Circularity Check

1 steps flagged · score 6.0 of 10

Target-adaptation pseudo-labels are generated from the same target data on which final AUROC is measured, so the reported gains partly measure the model's agreement with its own confident detections; a minor self-citation is not load-bearing.

  1. fitted input called prediction [Section 4.3 (Confident Detection Guided Adaptation), Section 5.1 (Experimental Setup), and Algorithm 2]
    "The top 𝑁𝑐𝑜𝑛 abnormal and normal detection pairs with the lowest entropy are selected as confident detections. After that, we use them as pseudo-labels for their corresponding edges. ... Then, we use the updated model for final anomaly detection on target test datasets."

    Section 5.1 states target fine-tuning uses "only normal edges and no abnormal ones," yet Section 4.3 and Algorithm 2 select abnormal and normal detection pairs as pseudo-labels on target data. Since abnormal detections cannot arise from a normal-only set, the only coherent reading is that pseudo-labeling runs on the same target data that, after anomaly injection, is also the test set. The alignment loss L_A then updates normal/abnormal prototypes to match these pseudo-labels, and the final anomaly score s_i = s_a,i - s_n,i is computed from those updated prototypes.

full rationale

The central claim is an empirical performance result, not a formal derivation, so most of the method is self-contained: source pretraining on Wikipedia and MOOC, temporal ego-graph extraction, dynamic prototype memory-buffer updates, and the BCE/alignment losses do not reduce to their inputs. There is minor self-citation (Ref. [6], co-authored by J. Liu) for distribution-based anomaly scoring, but it is a design choice, not a load-bearing uniqueness theorem, so it does not raise the score by itself. The significant circularity is in the target-adaptation protocol: Section 5.1 says target fine-tuning uses only normal edges, while Section 4.3 and Algorithm 2 select "abnormal and normal detection pairs" as pseudo-labels on target data. Because such pairs cannot exist in a normal-only set, the only consistent reading is that pseudo-labeling is applied to the same dataset that later serves as the test set with injected anomalies. The alignment loss then fits the prototypes to these pseudo-labels, and the final anomaly score is computed from those fitted prototypes; thus, for pseudo-labeled test edges, the prediction is an optimized reproduction of the model's own pre-adaptation confident detections, not an independent generalization. Appendix A.1 compounds this by injecting anomalies as links between "original disconnected nodes," making anomalous test edges structurally novel and easy to flag once the model has seen the test graph. As written, the text does not specify a temporal or random split separating adaptation edges from test edges, so the Table 2 gains are partly self-consistency rather than evidence for the generalist-DGAD claim. This warrants a score of 6: partial circularity via a prediction that is fitted to test-derived pseudo-labels.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central claim depends on several hand-chosen hyperparameters (M, N_con, alpha, lambda values) that are tuned on the target test sets, and on domain assumptions about ego-graph representativeness and pseudo-label reliability. No new physical or mathematical entities are introduced.

free parameters (7)
  • Memory buffer size M = 10% of training data size
    Chosen by hyperparameter search (Fig. 4) on target datasets; impacts number of prototypes and anomaly scoring.
  • Number of confident detections N_con = 10% of data size
    Tuned in Fig. 4; controls pseudo-label quality and adaptation.
  • Momentum alpha = 0.9
    Chosen in hyperparameter analysis; paper claims stability across values.
  • Loss ratio lambda_A = 0.1
    Tuned in Appendix C Table 5; balances alignment and BCE losses.
  • Loss ratio lambda_BCE = 0.9
    Tuned in Appendix C Table 5.
  • Ranking weights lambda_e and lambda_d = 0.7 and 0.3
    Tuned in Appendix C Table 6; control trade-off between domain similarity and prototype distinctiveness.
  • Learnable scoring weights lambda_n and lambda_a = learned, values not reported
    Introduced in Eqs. 15-16 with no initialization or range; their values are not reported.
assumptions (6)
  • domain assumption Temporal ego-graphs capture the anomalous pattern of an edge.
    Section 4.1 states 'for each edge, its anomalous pattern can be captured by ego-graph consisting of its neighbors' without proof; the entire representation hinges on this.
  • domain assumption Prototypes can be summarized by a small memory buffer of size M.
    Section 4.1 assumes the buffer retains 'the most discriminative prototypes' and that this is sufficient for cross-domain scoring.
  • domain assumption Low-entropy pseudo-labels are reliable enough for self-supervised adaptation.
    Section 4.3 selects top N_con low-entropy detections as pseudo-labels; if these are noisy, the adaptation can reinforce errors.
  • ad hoc to paper The anomaly score s_i = z^T mu - lambda z^T Sigma z is a valid distribution-based anomaly score.
    Eqs. 15-16 are presented without derivation and are not equivalent to a Mahalanobis distance; the scoring function is central to detection.
  • ad hoc to paper The covariance update in Eq. (13) is correct as written.
    Eq. (13) appears to be missing the momentum factor alpha on the previous covariance, inconsistent with Eq. (14); if the typo is in the paper, the actual implementation may differ.
  • domain assumption Euclidean distance between prototype pairs measures how discriminative they are.
    Eq. (6) defines difference score as mean Euclidean distance; no justification that this captures anomaly discriminability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DP-DGAD: A Generalist Dynamic Graph Anomaly Detector with Dynamic Prototypes." pith.science (2026). https://pith.science/paper/LUP3PLPO

@misc{pith2026250800664,
  author       = {Pith},
  title        = {Pith review of: DP-DGAD: A Generalist Dynamic Graph Anomaly Detector with Dynamic Prototypes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LUP3PLPO}},
  note         = {Machine review of arXiv:2508.00664}
}
read the original abstract

Dynamic graph anomaly detection (DGAD) is essential for identifying anomalies in evolving graphs across domains such as finance, traffic, and social networks. Recently, generalist graph anomaly detection (GAD) models have shown promising results. They are pretrained on multiple source datasets and generalize across domains. While effective on static graphs, they struggle to capture evolving anomalies in dynamic graphs. Moreover, the continuous emergence of new domains and the lack of labeled data further challenge generalist DGAD. Effective cross-domain DGAD requires both domain-specific and domain-agnostic anomalous patterns. Importantly, these patterns evolve temporally within and across domains. Building on these insights, we propose a DGAD model with Dynamic Prototypes (DP) to capture evolving domain-specific and domain-agnostic patterns. Firstly, DP-DGAD extracts dynamic prototypes, i.e., evolving representations of normal and anomalous patterns, from temporal ego-graphs and stores them in a memory buffer. The buffer is selectively updated to retain general, domain-agnostic patterns while incorporating new domain-specific ones. Then, an anomaly scorer compares incoming data with dynamic prototypes to flag both general and domain-specific anomalies. Finally, DP-DGAD employs confidence-based pseudo-labeling for effective self-supervised adaptation in target domains. Extensive experiments demonstrate state-of-the-art performance across ten real-world datasets from different domains.

Figures

Figures reproduced from arXiv: 2508.00664 by the authors.

Figure 1
Figure 1. (a) Anomalous patterns change over time. (b) Ex [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. General framework of DP-DGAD. (a) Starting with the first source dataset, we extract ego-graphs to capture temporal [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Model performance under continuous domain shift. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Impact of dynamic prototype memory buffer size [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Visualization of prototypes in memory buffer and [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 45 canonical work pages

  1. [1]

    Lei Cai, Zhengzhang Chen, Chen Luo, Jiaping Gui, Jingchao Ni, Ding Li, and Haifeng Chen. 2021. Structural temporal graph neural networks for anomaly de- tection in dynamic graphs. InProceedings of the 30th ACM international conference on Information & Knowledge Management . 3747–3756

    work page 2021

  2. [2]

    João Carvalho, Mengtao Zhang, Robin Geyer, Carlos Cotrini, and Joachim M Buhmann. 2023. Invariant anomaly detection under distribution shifts: a causal perspective. Advances in Neural Information Processing Systems 36 (2023), 56310– 56337

    work page 2023

  3. [3]

    Dong Chen, Xiang Zhao, and Weidong Xiao. 2024. Fine-grained anomaly detec- tion on dynamic graphs via attention alignment. In 2024 IEEE 40th International Conference on Data Engineering (ICDE) . IEEE, 3178–3190

    work page 2024

  4. [4]

    Zining Chen, Xingshuang Luo, Weiqiu Wang, Zhicheng Zhao, Fei Su, and Aidong Men. 2025. Filter or compensate: Towards invariant representation from dis- tribution shift for anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 2420–2428

    work page 2025

  5. [5]

    Leyan Deng, Defu Lian, Zhenya Huang, and Enhong Chen. 2022. Graph con- volutional adversarial networks for spatiotemporal anomaly detection. IEEE Transactions on Neural Networks and Learning Systems 33, 6 (2022), 2416–2428

    work page 2022

  6. [6]

    Xiaoqing Guo, Jie Liu, and Yixuan Yuan. 2024. Infproto-Powered Adaptive Classifier and Agnostic Feature Learning for Single Domain Generalization in Medical Images. International Journal of Computer Vision 132, 12 (2024), 5905– 5928

    work page 2024

  7. [7]

    Xingzhi Guo, Baojian Zhou, and Steven Skiena. 2022. Subset node anomaly tracking over large dynamic graphs. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 475–485

    work page 2022

  8. [8]

    Zhongxuan Han, Xiaolin Zheng, Chaochao Chen, Wenjie Cheng, and Yang Yao

    work page

Show all 50 references
  1. [9]

    Lingbai Kong, Wengen Li, Hanchen Yang, Yichao Zhang, Jihong Guan, and Shuigeng Zhou. 2024. Causalformer: An interpretable transformer for temporal causal discovery. IEEE Transactions on Knowledge and Data Engineering (2024)

    2024

  2. [10]

    Srijan Kumar, Francesca Spezzano, VS Subrahmanian, and Christos Faloutsos

  3. [11]

    Srijan Kumar, Xikun Zhang, and Jure Leskovec. 2019. Predicting dynamic em- bedding trajectory in temporal interaction networks. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining . 1269–1278

    2019

  4. [12]

    Jongha Lee, Sunwoo Kim, and Kijung Shin. 2024. Slade: Detecting dynamic anom- alies in edge streams without labels via self-supervised learning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 1506–1517

    2024

  5. [13]

    Aodong Li, Chen Qiu, Marius Kloft, Padhraic Smyth, Maja Rudolph, and Stephan Mandt. 2023. Zero-shot anomaly detection via batch normalization. Advances in Neural Information Processing Systems 36 (2023), 40963–40993

    2023

  6. [14]

    Yixin Liu, Shiyuan Li, Yu Zheng, Qingfeng Chen, Chengqi Zhang, and Shirui Pan. 2024. ARC: a generalist graph anomaly detector with in-context learning. In Proceedings of the 38th International Conference on Neural Information Processing Systems. 50772–50804

    2024

  7. [15]

    Yixin Liu, Shirui Pan, Yu Guang Wang, Fei Xiong, Liang Wang, Qingfeng Chen, and Vincent CS Lee. 2021. Anomaly detection in dynamic graphs via transformer. IEEE Transactions on Knowledge and Data Engineering 35, 12 (2021), 12081–12094

    2021

  8. [16]

    Zhe Liu, Xiang Huang, Jingyun Zhang, Zhifeng Hao, Li Sun, and Hao Peng. 2024. Multivariate time-series anomaly detection based on enhancing graph attention networks with topological analysis. In Proceedings of the 33rd ACM International Conference on Information and Knowledge ...

    2024

  9. [17]

    Zemin Liu, Xingtong Yu, Yuan Fang, and Xinming Zhang. 2023. Graphprompt: Unifying pre-training and downstream tasks for graph neural networks. In Pro- ceedings of the ACM web conference 2023 . 417–428

    2023

  10. [18]

    Xiaoxiao Ma, Ruikun Li, Fanzhen Liu, Kaize Ding, Jian Yang, and Jia Wu. 2024. Graph anomaly detection with few labels: A data-centric approach. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 2153–2164

    2024

  11. [19]

    Chaoxi Niu, Hezhe Qiao, Changlu Chen, Ling Chen, and Guansong Pang

  12. [20]

    Tore Opsahl and Pietro Panzarasa. 2009. Clustering in weighted networks. Social networks 31, 2 (2009), 155–163

    2009

  13. [21]

    Junjun Pan, Yixin Liu, Xin Zheng, Yizhen Zheng, Alan Wee-Chung Liew, Fuyi Li, and Shirui Pan. 2025. A label-free heterophily-guided approach for unsuper- vised graph fraud detection. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 12443–12451

    2025

  14. [22]

    Hezhe Qiao, Chaoxi Niu, Ling Chen, and Guansong Pang. 2025. AnomalyGFM: Graph foundation model for zero/few-shot anomaly detection. arXiv preprint arXiv:2502.09254 (2025)

    2025 arXiv

  15. [23]

    Ziyue Qiao, Qianyi Cai, Hao Dong, Jiawei Gu, Pengyang Wang, Meng Xiao, Xiao Luo, and Hui Xiong. 2025. GCAL: Adapting Graph Models to Evolving Domain Shifts. In Forty-second International Conference on Machine Learning

    2025

  16. [24]

    Ryan Rossi and Nesreen Ahmed. 2015. The network data repository with inter- active graph analytics and visualization. In Proceedings of the AAAI conference on artificial intelligence, Vol. 29

    2015

  17. [25]

    Haopeng Sun, Lumin Xu, Sheng Jin, Ping Luo, Chen Qian, and Wentao Liu

  18. [26]

    Sheng Tian, Jihai Dong, Jintang Li, Wenlong Zhao, Xiaolong Xu, Bowen Song, Changhua Meng, Tianyi Zhang, Liang Chen, et al. 2023. Sad: Semi-supervised anomaly detection on dynamic graphs. arXiv preprint arXiv:2305.13573 (2023)

    2023 arXiv

  19. [27]

    Yuntong Tian, Jiaxi Li, Huazhu Fu, Lei Zhu, Lequan Yu, and Liang Wan. 2024. Self- mining the confident prototypes for source-free unsupervised domain adaptation in image segmentation. IEEE Transactions on Multimedia 26 (2024), 7709–7720

    2024

  20. [28]

    Yuxing Tian, Yiyan Qi, and Fan Guo. 2024. Freedyg: Frequency enhanced continuous-time dynamic graph model for link prediction. In The twelfth in- ternational conference on learning representations

    2024

  21. [29]

    In The twelfth international conference on learning representations

    Program: Prototype graph model based pseudo-label learning for test-time adaptation. In The twelfth international conference on learning representations

  22. [30]

    Huan Wang, Junyang Chen, Yirui Wu, Victor CM Leung, and Di Wang. 2025. Epm: Evolutionary perception method for anomaly detection in noisy dynamic graphs. IEEE Transactions on Knowledge and Data Engineering (2025)

    2025

  23. [31]

    Hongjun Wang and Yisen Wang. 2023. Generalist: Decoupling natural and robust generalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 20554–20563

    2023

  24. [32]

    Yuwen Wang, Shunyu Liu, Tongya Zheng, Kaixuan Chen, and Mingli Song. 2024. Unveiling global interactive patterns across graphs: Towards interpretable graph neural networks. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 3277–3288

    2024

  25. [33]

    Guancheng Wan, Wenke Huang, and Mang Ye. 2024. Federated graph learning under domain shift with generalizable prototypes. In Proceedings of the AAAI conference on artificial intelligence , Vol. 38. 15429–15437. Conference KDD ’XX, June 03–05, 2018, Woodstock, NY Zheng et al

    2024

  26. [34]

    Jiacong Xu, Shao-Yuan Lo, Bardia Safaei, Vishal M Patel, and Isht Dwivedi. 2025. Towards zero-shot anomaly detection and reasoning with multimodal large language models. In Proceedings of the Computer Vision and Pattern Recognition Conference. 20370–20382

    2025

  27. [35]

    Yiming Xu, Bin Shi, Teng Ma, Bo Dong, Haoyi Zhou, and Qinghua Zheng. 2023. Cldg: Contrastive learning on dynamic graphs. In 2023 IEEE 39th International Conference on Data Engineering (ICDE) . IEEE, 696–707

    2023

  28. [36]

    Hanchen Yang, Jiannong Cao, Wengen Li, Shuyu Wang, Hui Li, Jihong Guan, and Shuigeng Zhou. 2025. Spatial-Temporal Data Mining for Ocean Science: Data, Methodologies and Opportunities. ACM Transactions on Knowledge Discovery from Data (July 2025), 1–46

    2025

  29. [37]

    Yuxia Wu, Lizi Liao, and Yuan Fang. 2025. Retrieval augmented generation for dynamic graph modeling. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval . 1434–1443

    2025

  30. [38]

    Hanchen Yang, Wengen Li, Siyun Hou, Jihong Guan, and Shuigeng Zhou. 2023. HiGRN: a hierarchical graph recurrent network for global sea surface temperature prediction. ACM Transactions on Intelligent Systems and Technology 14, 4 (2023), 1–19

    2023

  31. [39]

    Jaewon Yang and Jure Leskovec. 2012. Defining and evaluating network commu- nities based on ground-truth. In Proceedings of the ACM SIGKDD workshop on mining data semantics. 1–8

    2012

  32. [40]

    Xiao Yang, Xuejiao Zhao, and Zhiqi Shen. 2025. A generalizable anomaly detection method in dynamic graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 22001–22009

    2025

  33. [41]

    Hanchen Yang, Yuwei Chen, Jiannong Cao, Wengen Li, Yu Yang, Shuyu Wang, Jihong Guan, Rufu Qin, and Shuigeng Zhou. 2024. UniOcean: A Unified Frame- work for Predicting Multiple Ocean Factors of Varying Temporal Scales. IEEE Transactions on Geoscience and Remote Sensing (2024)

    2024

  34. [42]

    Haihong Zhao, Aochuan Chen, Xiangguo Sun, Hong Cheng, and Jia Li. 2024. All in one and one for all: A simple yet effective method towards cross-domain graph pretraining. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 4443–4454

    2024

  35. [43]

    Jialun Zheng, Divya Saxena, Jiannong Cao, Hanchen Yang, and Penghui Ruan

  36. [44]

    Jiawen Zhu and Guansong Pang. 2024. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 17826–17836

    2024

  37. [45]

    Beichuan Zhang, Raymond Liu, Daniel Massey, and Lixia Zhang. 2005. Collecting the Internet AS-level topology.ACM SIGCOMM Computer Communication Review 35, 1 (2005), 53–61

    2005

  38. [48]

    In International Conference on Database Systems for Advanced Applications

    Inductive Spatial Temporal Prediction Under Data Drift with Informative Graph Neural Network. In International Conference on Database Systems for Advanced Applications. Springer, 169–185

  39. [50]

    hijacked

    Zulun Zhu, Kai Wang, Haoyu Liu, Jintang Li, and Siqiang Luo. 2024. Topology- monitorable contrastive learning on dynamic graphs. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 4700–4711. A DETAILED EXPERIMENTS DESCRIPTION A.1 Datasets...

    2024

  40. [2016]

    In 2016 IEEE 16th international conference on data mining (ICDM)

    Edge weight prediction in weighted signed networks. In 2016 IEEE 16th international conference on data mining (ICDM) . IEEE, 221–230

    2016

  41. [2023]

    In Proceedings of the ACM Web Conference 2023

    Intra and inter domain hypergraph convolutional network for cross-domain recommendation. In Proceedings of the ACM Web Conference 2023 . 449–459

    2023

  42. [2024]

    arXiv preprint arXiv:2410.14886 (2024)

    Zero-shot generalist graph anomaly detection with unified neighborhood prompts. arXiv preprint arXiv:2410.14886 (2024)

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.