Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

Out-of-Distribution Detection on Graphs: A Survey

T0 review · 2 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This survey proposes a unified formal definition of graph out-of-distribution detection and a four-family taxonomy of existing methods.

desk verdict A useful first map of graph OOD detection, but its 'rigorous definition' has a fixable gap that should be addressed before it becomes the field's standard reference. read the letter →

arxiv 2502.08105 v1 pith:6IQ4TFSY submitted 2025-02-12 cs.LG

Tingyi Cai , Yunliang Jiang , Yixin Liu , Ming Li , Changqin Huang , Shirui Pan This is my paper · ORCID
classification cs.LG
keywords graphout-of-distributiondetectionneuralnetworkstaxonomydistributionshiftopen-worldlearninganomalybenchmark
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

This paper aims to give graph out-of-distribution (GOOD) detection a single formal definition and a systematic taxonomy of solution strategies. The survey argues that GOOD detection is distinct from graph anomaly detection, outlier detection, and GOOD generalization, and that essentially all existing methods can be grouped into four families: enhancement-based, reconstruction-based, information propagation-based, and classification-based. By organizing more than thirty published methods into these families and connecting them to standard datasets and benchmarks, the paper tries to turn a fragmented research area into a field with shared vocabulary. If the framing holds, future work can state its contribution by saying which family it extends, and evaluators can compare methods on common ground.

What carries the argument

The load-bearing device is the pair formed by the formal task definition (Definition 1) and the two-term loss (Definition 2), plus the four-way taxonomy that hangs off them. Enhancement-based methods widen or improve the training setup; reconstruction-based methods use generative models trained on in-distribution data so that out-of-distribution samples yield poor reconstructions; information propagation-based methods rely on message passing to spread features, energies, or uncertainties; classification-based methods use probabilistic evidence or boundary optimization to flag low-confidence samples. This structure is what lets more than thirty methods be compared as alternative implementations of the same rejection rule.

What would settle it

A single published GOOD detection method whose scoring mechanism uses none of the four families, meaning no enhancement of data or model, no reconstruction error, no message-passing propagation, and no classifier-based uncertainty, would refute the survey's claim that the taxonomy is comprehensive.

Watch

Extended reading notes

Core claim

The paper's central claim is that every GOOD detection method solves one shared decision problem: given a node or graph sample, accept it when it comes from the known in-distribution $D_{\mathrm{in}}$ and reject it when it comes from the unknown out-distribution $D_{\mathrm{out}}$, guided by a loss $L_{\mathrm{GOOD}}$ that balances classification accuracy on in-distribution samples against rejection of out-of-distribution samples. The survey then claims to show a systematic four-way split in how methods implement that rule: enhancing the data or model, reconstructing the input and scoring the reconstruction error, propagating feature/energy/uncertainty signals along the graph, or reading OOD scores from classifier outputs.

Load-bearing premise

The formal loss assumes the model can take an expectation over out-of-distribution samples, but those samples are by definition unknown and unavailable at training time, so the objective is only computable if a method generates its own synthetic outliers or uses exposure data.

Editorial extensions

If this is right

  • Every future GOOD detection paper can position itself as an extension of one of the four families, which lowers the cost of comparing new methods.
  • The separation from graph anomaly detection implies that evaluations should use distribution-shift partitions, such as class-based, domain-based, or perturbed-structure splits, rather than only anomalous-sample lists.
  • Reconstruction-based and enhancement-based methods can operate without labeled OOD data, which matters for applications where OOD examples are rare or undefined.
  • The DrugOOD and UB-GOLD benchmarks give the field two concrete points of comparison for graph-level GOOD detection.

Reading between the lines

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

  • Because Definition 2 requires an expectation over $D_{\mathrm{out}}$ while Definition 1 declares $D_{\mathrm{out}}$ unknown, any practical method must synthesize or expose its own pseudo-OOD data; the paper's data-enhancement family can be read as supplying exactly that missing training signal.
  • The four-way taxonomy is more likely a set of design dimensions than a strict partition; a method that jointly enhances data, reconstructs input, propagates information, and classifies would blur all four boundaries, and testing such a hybrid would stress the taxonomy.
  • The survey's stress points, such as heterophilic graphs, temporal shifts, and noisy structures, suggest a natural next benchmark: one that varies homophily and time while holding the four method families fixed, to see which family's assumptions break first.
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

2 major / 6 minor

Summary. The paper is a survey of out-of-distribution (OOD) detection on graphs, which it calls GOOD detection. It proposes a formal definition of the task (Definition 1), a composite training loss (Definition 2), and a taxonomy that splits existing methods into four families: enhancement-based, reconstruction-based, information propagation-based, and classification-based approaches. The survey reviews more than 30 works from 2020–2025, organizes them within this taxonomy, discusses benchmark datasets and evaluation protocols, distinguishes GOOD detection from graph anomaly detection, outlier detection, and GOOD generalization, and outlines future research directions.

Significance. If its definitions and taxonomy hold up, the survey provides a a useful shared vocabulary for a rapidly growing subfield and a convenient entry point for researchers. It is broad in coverage, includes a public repository, and the subdivision of information propagation methods into feature-, energy-, and uncertainty-propagation is a practical heuristic. The paper also makes a credible effort to separate GOOD detection from adjacent problems such as anomaly detection and GOOD generalization. However, the paper's own stated contribution of a 'rigorous definition' is weakened by a gap in the formalization: the training objective in Definition 2 depends on samples from Dout, which Definition 1 declares unavailable. Because this gap affects a core claim of the paper, the definition needs repair before the survey can serve as a reference for the field's formal foundations.

major comments (2)
  1. [Section 2.1, Definition 2, Eq. (2)] The second expectation in Eq. (2) is taken over samples from Dout, but Definition 1 defines Dout as an unknown distribution whose samples are encountered only at test time and are to be rejected. Under that definition, no Dout samples are available to compute the expectation during training. The paper does not state the implicit assumption that this term is computed with surrogate OOD data, pseudo-OOD examples, or an auxiliary outlier-exposure set; in fact, several surveyed methods (e.g., GOOD-D, GOLD, SMUG) explicitly avoid real OOD samples. Please either reformulate Eq. (2) as a conceptual objective that uses a surrogate distribution (and state this explicitly, with the relationship to the surrogate-generation techniques in Sections 3 and 6), or discuss the necessary assumptions. Without such a repair, the claimed 'rigorous definition' does not formalize the actual optimization problem faced by the surveyed field.
  2. [Section 2.2 and Figure 2] The four taxonomy categories are not defined with exclusion criteria or assignment rules, so a method can plausibly fall into more than one category. For example, energy-propagation methods (e.g., GNNSafe, NODESAFE) derive scores from a classifier's logits, blurring the line with classification-based methods, while reconstruction-based methods often feed reconstruction errors into a classifier or thresholding rule. The paper should state the criterion used for placing a method into exactly one category (e.g., the primary mechanism or the main training objective) or explicitly acknowledge that categories overlap and provide guidance for resolving ambiguous cases. Without such criteria, the taxonomy is a loose grouping rather than a 'systematic categorization' as claimed in the abstract.
minor comments (6)
  1. [Author contact information] The email for Shirui Pan is written as 's.pan@griffth.edu.au', which appears to be a typo for 's.pan@griffith.edu.au'.
  2. [Table 1] The table caption says 'highlighted multi-graph datasets', but no visual highlighting is present in the rendered table; the text should describe which datasets are multi-graph and flag them explicitly.
  3. [Table 2] The caption refers to 'blue highlighted node-level datasets', but no blue highlighting appears in the table text; this needs to be described or rendered.
  4. [Section 2.1, Eq. (1)] The notation uses 'reject' as a value of y-hat, which is not a class label; please clarify that rejection is a decision output (e.g., based on a threshold on a score) rather than a label.
  5. [Section 8.2] The distinction between GOOD detection (semantic shift) and GOOD generalization (covariate shift) is asserted but not reconciled with Definition 1, which does not restrict Dout to semantic shifts; please clarify how the formal definition maps to this dichotomy.
  6. [Section 8.3] The text cites GRADATE and BWGNN as applications of GOOD detection, but these methods are typically presented as anomaly detection; since Section 8.1 argues for distinctions, a brief justification of why they qualify as GOOD detection would avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the survey organizes and reviews existing methods by describing them, but it makes no derivation or data-fitting claim whose output is presupposed by its input.

full rationale

This paper is a survey, not a method paper with a derivation chain. Its central claims are taxonomic and descriptive: it defines GOOD detection, proposes a four-way categorization of existing methods, and reviews the literature. There is no fitted parameter that is later renamed as a prediction, no theorem whose proof assumes its own conclusion, and no quantitative result that is forced by construction. The authors do cite several of their own prior works (e.g., GOOD-D, GOODAT, PGR-MOOD, ML-GOOD, OpenWGL, and the UB-GOLD benchmark), but these are primary sources of the methods being surveyed; they are not invoked as external evidence to justify the taxonomy. The taxonomy itself is presented as an organizing structure, not as a result derived from those papers. The only formal object is Definition 2, which writes the GOOD detection loss with an expectation over Dout even though Definition 1 defines Dout as an unknown test-time distribution. That is a formalization weakness or a correctness concern about whether the objective is computable as stated, not a circularity: the definition is not used to prove a later claim by assuming it. Accordingly, no circular step is identified, and the appropriate score is 0.

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

The survey's central contributions are organizational, so the load-bearing assumptions are about the taxonomy's completeness and the formalization's computability, not about fitted parameters or new physical entities. The main issue is that Eq. (2) assumes access to Dout, which is an unrecognized assumption.

assumptions (3)
  • domain assumption The four-way taxonomy (enhancement, reconstruction, information propagation, classification) is exhaustive and mutually exclusive.
    The survey assigns every reviewed method to exactly one of four categories but provides no rule for overlap; for example, energy-based methods use a GNN classifier and propagation, so the categories are not provably disjoint. See Section 2.2 and Figure 2.
  • ad hoc to paper The optimization objective in Definition 2 can include samples from Dout during training.
    Eq. (2) defines LGOOD with an expectation over Dout, which is an unknown distribution by Definition 1. This is not stated as an assumption, and it is not true in the standard unsupervised detection setup. Location: Section 2.1, Eq. (2).
  • domain assumption Standard OOD detection notations and definitions (Din/Dout, reject option) transfer unchanged from image/text domains to graph data.
    Definition 1 imports the standard OOD formulation without discussing graph-specific components such as the adjacency matrix A in the sample space. The paper later discusses graph-specific challenges, but the formal definition itself does not encode them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Out-of-Distribution Detection on Graphs: A Survey." pith.science (2026). https://pith.science/paper/6IQ4TFSY

@misc{pith2026250208105,
  author       = {Pith},
  title        = {Pith review of: Out-of-Distribution Detection on Graphs: A Survey},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6IQ4TFSY}},
  note         = {Machine review of arXiv:2502.08105}
}
read the original abstract

Graph machine learning has witnessed rapid growth, driving advancements across diverse domains. However, the in-distribution assumption, where training and testing data share the same distribution, often breaks in real-world scenarios, leading to degraded model performance under distribution shifts. This challenge has catalyzed interest in graph out-of-distribution (GOOD) detection, which focuses on identifying graph data that deviates from the distribution seen during training, thereby enhancing model robustness. In this paper, we provide a rigorous definition of GOOD detection and systematically categorize existing methods into four types: enhancement-based, reconstruction-based, information propagation-based, and classification-based approaches. We analyze the principles and mechanisms of each approach and clarify the distinctions between GOOD detection and related fields, such as graph anomaly detection, outlier detection, and GOOD generalization. Beyond methodology, we discuss practical applications and theoretical foundations, highlighting the unique challenges posed by graph data. Finally, we discuss the primary challenges and propose future directions to advance this emerging field. The repository of this survey is available at https://github.com/ca1man-2022/Awesome-GOOD-Detection.

Figures

Figures reproduced from arXiv: 2502.08105 by the authors.

Figure 1
Figure 1. Applications and rising interest in GOOD detection. This [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. These methods incorporate advanced techniques [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figure 2
Figure 2. A taxonomy of GOOD detection approaches. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figures from the paper (3 more)
Figure 3
Figure 3. Figure 3: General framework for the enhancement-based methods. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png]
Figure 4
Figure 4. Figure 4: General framework for the reconstruction-based methods. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: General framework for the classification-based methods. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Both Topology and Text Matter: Revisiting LLM-guided Out-of-Distribution Detection on Text-attributed Graphs

    cs.LG 2026-02 conditional novelty 6.0 of 10

    LG-Plug mines pseudo-OOD exposures from clustered unlabeled nodes via iterative LLM prompting and regularizes topology-driven graph OOD detectors, cutting FPR95 by ≥7% across six TAG benchmarks.

Reference graph

Works this paper leans on

46 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    Graph out-of- distribution detection goes neighborhood shaping

    [Bao et al., 2024] Tianyi Bao, Qitian Wu, Zetian Jiang, Yit- ing Chen, Jiawei Sun, and Junchi Yan. Graph out-of- distribution detection goes neighborhood shaping. In ICML,

    work page 2024

  2. [6]

    SGOOD: Substructure-enhanced graph-level out- of-distribution detection

    [Ding et al., 2024] Zhihao Ding, Jieming Shi, Shiqi Shen, Xuequn Shang, Jiannong Cao, Zhipeng Wang, and Zhi Gong. SGOOD: Substructure-enhanced graph-level out- of-distribution detection. In CIKM, pages 467–476,

    work page 2024

  3. [7]

    When and how does in-distribution label help out-of- distribution detection? In ICML,

    [Du et al., 2024] Xuefeng Du, Yiyou Sun, and Yixuan Li. When and how does in-distribution label help out-of- distribution detection? In ICML,

    work page 2024

  4. [8]

    Graph anomaly detection via multi-scale contrastive learn- ing networks with augmented view

    [Duan et al., 2023] Jingcan Duan, Siwei Wang, Pei Zhang, En Zhu, Jingtao Hu, Hu Jin, Yue Liu, and Zhibin Dong. Graph anomaly detection via multi-scale contrastive learn- ing networks with augmented view. In AAAI, pages 7459– 7467,

    work page 2023

  5. [9]

    Al- leviating structural distribution shift in graph anomaly de- tection

    [Gao et al., 2023] Yuan Gao, Xiang Wang, Xiangnan He, Zhenguang Liu, Huamin Feng, and Yongdong Zhang. Al- leviating structural distribution shift in graph anomaly de- tection. In WSDM, pages 357–365,

    work page 2023

  6. [10]

    An energy-centric framework for category-free out-of- distribution node detection in graphs

    [Gong and Sun, 2024] Zheng Gong and Ying Sun. An energy-centric framework for category-free out-of- distribution node detection in graphs. In KDD, page 908–919,

    work page 2024

  7. [11]

    GOOD: A graph out-of-distribution bench- mark

    [Gui et al., 2022] Shurui Gui, Xiner Li, Limei Wang, and Shuiwang Ji. GOOD: A graph out-of-distribution bench- mark. In NeurIPS, pages 2059–2073,

    work page 2022

  8. [13]

    Improvements on uncertainty quantification for node classification via distance based regularization

    [Hart et al., 2023] Russell Hart, Linlin Yu, Yifei Lou, and Feng Chen. Improvements on uncertainty quantification for node classification via distance based regularization. In NeurIPS, pages 55454–55478,

    work page 2023

Show all 46 references
  1. [14]

    Open-world lifelong graph learning

    [Hoffmann et al., 2023] Marcel Hoffmann, Lukas Galke, and Ansgar Scherp. Open-world lifelong graph learning. In IJCNN,

    2023

  2. [15]

    Open graph benchmark: Datasets for machine learning on graphs

    [Hu et al., 2020] Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. In NeurIPS, pages 22118– 22133,

    2020

  3. [17]

    DrugOOD: Out-of- distribution dataset curator and benchmark for ai-aided drug discovery–a focus on affinity prediction problems with noise annotations

    [Ji et al., 2023] Yuanfeng Ji, Lu Zhang, Jiaxiang Wu, Bingzhe Wu, Lanqing Li, Long-Kai Huang, Tingyang Xu, Yu Rong, Jie Ren, Ding Xue, et al. DrugOOD: Out-of- distribution dataset curator and benchmark for ai-aided drug discovery–a focus on affinity prediction problems with no...

    2023

  4. [18]

    A survey of graph neural net- works in real world: Imbalance, noise, privacy and OOD challenges

    [Ju et al., 2024] Wei Ju, Siyu Yi, Yifan Wang, Zhiping Xiao, Zhengyang Mao, Hourun Li, Yiyang Gu, Yifang Qin, Nan Yin, Senzhang Wang, et al. A survey of graph neural net- works in real world: Imbalance, noise, privacy and OOD challenges. arXiv preprint arXiv:2403.04468,

    2024

  5. [19]

    HGOE: Hybrid external and internal graph outlier expo- sure for graph out-of-distribution detection

    [Junwei et al., 2024] He Junwei, Qianqian Xu, Yangbangyan Jiang, Zitai Wang, Yuchen Sun, and Qingming Huang. HGOE: Hybrid external and internal graph outlier expo- sure for graph out-of-distribution detection. In MM, pages 1544–1553,

    2024

  6. [20]

    Neural relation graph: A unified frame- work for identifying label noise and outlier data

    [Kim et al., 2023] Jang-Hyun Kim, Sangdoo Yun, and Hyun Oh Song. Neural relation graph: A unified frame- work for identifying label noise and outlier data. In NeurIPS, pages 43754–43779,

    2023

  7. [21]

    Predicting dynamic embedding trajectory in temporal interaction networks

    [Kumar et al., 2019] Srijan Kumar, Xikun Zhang, and Jure Leskovec. Predicting dynamic embedding trajectory in temporal interaction networks. InKDD, pages 1269–1278,

    2019

  8. [23]

    Graph neural stochastic diffusion for estimating uncertainty in node classification

    [Lin et al., 2024] Xixun Lin, Wenxiao Zhang, Fengzhao Shi, Chuan Zhou, Lixin Zou, Xiangyu Zhao, Dawei Yin, Shirui Pan, and Yanan Cao. Graph neural stochastic diffusion for estimating uncertainty in node classification. In ICML,

    2024

  9. [24]

    GOOD-D: On unsupervised graph out-of- distribution detection

    [Liu et al., 2023] Yixin Liu, Kaize Ding, Huan Liu, and Shirui Pan. GOOD-D: On unsupervised graph out-of- distribution detection. In WSDM, pages 339–347,

    2023

  10. [25]

    Towards semi-supervised univer- sal graph classification

    [Luo et al., 2023] Xiao Luo, Yusheng Zhao, Yifang Qin, Wei Ju, and Ming Zhang. Towards semi-supervised univer- sal graph classification. IEEE Transactions on Knowledge and Data Engineering, 36(1):416–428,

    2023

  11. [26]

    Revisiting score propagation in graph out-of-distribution detection

    [Ma et al., 2024] Longfei Ma, Yiyou Sun, Kaize Ding, Zemin Liu, and Fei Wu. Revisiting score propagation in graph out-of-distribution detection. In NeurIPS,

    2024

  12. [27]

    [McAuley et al., 2015] Julian McAuley, Rahul Pandey, and Jure Leskovec

    URL: https://openreview.net/forum?id=jb5qN3212b. [McAuley et al., 2015] Julian McAuley, Rahul Pandey, and Jure Leskovec. Inferring networks of substitutable and complementary products. In KDD, pages 785–794,

    2015

  13. [30]

    Collective classification in network data

    [Sen et al., 2008] Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi-Rad. Collective classification in network data. AI magazine , 29(3):93–93,

    2008

  14. [33]

    Calibrate graph neural networks under out-of-distribution nodes via deep Q-learning

    [Shi et al., 2023] Weili Shi, Xueying Yang, Xujiang Zhao, Haifeng Chen, Zhiqiang Tao, and Sheng Li. Calibrate graph neural networks under out-of-distribution nodes via deep Q-learning. In CIKM, pages 2270–2279,

    2023

  15. [34]

    Learn- ing on graphs with out-of-distribution nodes

    [Song and Wang, 2022] Yu Song and Donglin Wang. Learn- ing on graphs with out-of-distribution nodes. In KDD, pages 1635–1645,

    2022

  16. [35]

    Graph posterior network: Bayesian predic- tive uncertainty for node classification

    [Stadler et al., 2021] Maximilian Stadler, Bertrand Charp- entier, Simon Geisler, Daniel Z ¨ugner, and Stephan G¨unnemann. Graph posterior network: Bayesian predic- tive uncertainty for node classification. In NeurIPS, pages 18033–18048,

    2021

  17. [36]

    ArnetMiner: Extraction and mining of academic social networks

    [Tang et al., 2008] Jie Tang, Jing Zhang, Limin Yao, Juanzi Li, Li Zhang, and Zhong Su. ArnetMiner: Extraction and mining of academic social networks. In KDD, pages 990– 998,

    2008

  18. [37]

    Rethinking graph neural networks for anomaly de- tection

    [Tang et al., 2022] Jianheng Tang, Jiajin Li, Ziqi Gao, and Jia Li. Rethinking graph neural networks for anomaly de- tection. In ICML, pages 21076–21089,

    2022

  19. [38]

    Spreading out-of- distribution detection on graphs

    [Um et al., 2025] Daeho Um, Jongin Lim, Sunoh Kim, Yuneil Yeo, and Yoonho Jung. Spreading out-of- distribution detection on graphs. In ICLR,

    2025

  20. [39]

    Unifying unsupervised graph-level anomaly detection and out-of-distribution detection: A benchmark

    [Wang et al., 2024c] Yili Wang, Yixin Liu, Xu Shen, Chenyu Li, Kaize Ding, Rui Miao, Ying Wang, Shirui Pan, and Xin Wang. Unifying unsupervised graph-level anomaly detection and out-of-distribution detection: A benchmark. arXiv preprint arXiv:2406.15523,

    arXiv

  21. [40]

    GOLD: Graph out-of-distribution de- tection via implicit adversarial latent generation

    [Wang et al., 2025] Danny Wang, Ruihong Qiu, Guangdong Bai, and Zi Huang. GOLD: Graph out-of-distribution de- tection via implicit adversarial latent generation. In ICLR,

    2025

  22. [41]

    OpenWGL: Open-world graph learning

    [Wu et al., 2020] Man Wu, Shirui Pan, and Xingquan Zhu. OpenWGL: Open-world graph learning. In ICDM, pages 681–690,

    2020

  23. [42]

    Energy-based out-of-distribution detec- tion for graph neural networks

    [Wu et al., 2023] Qitian Wu, Yiting Chen, Chenxiao Yang, and Junchi Yan. Energy-based out-of-distribution detec- tion for graph neural networks. In ICLR,

    2023

  24. [43]

    [Zhang et al., 2022] Qin Zhang, Qincai Li, Xiaojun Chen, Peng Zhang, Shirui Pan, Philippe Fournier-Viger, and Joshua Zhexue Huang

    URL: https://openreview.net/forum?id=mjh7AOWozN. [Zhang et al., 2022] Qin Zhang, Qincai Li, Xiaojun Chen, Peng Zhang, Shirui Pan, Philippe Fournier-Viger, and Joshua Zhexue Huang. A dynamic variational framework for open-world node classification in structured sequences. In IC...

    2022

  25. [44]

    Uncertainty aware semi-supervised learning on graph data

    [Zhao et al., 2020] Xujiang Zhao, Feng Chen, Shu Hu, and Jin-Hee Cho. Uncertainty aware semi-supervised learning on graph data. In NeurIPS, pages 12827–12836,

    2020

  26. [45]

    Well- classified examples are underestimated in classification with deep neural networks

    [Zhao et al., 2022] Guangxiang Zhao, Wenkai Yang, Xu- ancheng Ren, Lei Li, Yunfang Wu, and Xu Sun. Well- classified examples are underestimated in classification with deep neural networks. In AAAI, pages 9180–9189,

    2022

  27. [46]

    FocusedCleaner: Sanitizing poi- soned graphs for robust GNN-based node classification

    [Zhu et al., 2024] Yulin Zhu, Liang Tong, Gaolei Li, Xiapu Luo, and Kai Zhou. FocusedCleaner: Sanitizing poi- soned graphs for robust GNN-based node classification. IEEE Transactions on Knowledge and Data Engineering , 36(6):2476–2489, 2024

    2024

  28. [2008]

    Pitfalls of graph neural network evaluation

    [Shchur et al., 2018] Oleksandr Shchur, Maximil- ian Mumme, Aleksandar Bojchevski, and Stephan G¨unnemann. Pitfalls of graph neural network evaluation. arXiv preprint arXiv:1811.05868,

    2018 arXiv

  29. [2015]

    Uncertainty-aware graph-based multimodal remote sensing detection of out-of-distribution samples

    [Rolland et al., 2021] Iain Rolland, Andrea Marinoni, and Sivasakthy Selvakumaran. Uncertainty-aware graph-based multimodal remote sensing detection of out-of-distribution samples. In CIKM Workshops,

    2021

  30. [2017]

    ML- GOOD: Towards multi-label graph out-of-distribution de- tection

    [Cai et al., 2025] Tingyi Cai, Yunliang Jiang, Ming Li, Changqin Huang, Yi Wang, and Qionghao Huang. ML- GOOD: Towards multi-label graph out-of-distribution de- tection. In AAAI,

    2025

  31. [2018]

    Optimizing OOD detection in molecular graphs: A novel approach with diffusion mod- els

    [Shen et al., 2024] Xu Shen, Yili Wang, Kaixiong Zhou, Shirui Pan, and Xin Wang. Optimizing OOD detection in molecular graphs: A novel approach with diffusion mod- els. KDD, pages 2640–2650,

    2024

  32. [2019]

    GraphDE: A generative framework for debiased learning and out-of-distribution detection on graphs

    [Li et al., 2022] Zenan Li, Qitian Wu, Fan Nie, and Junchi Yan. GraphDE: A generative framework for debiased learning and out-of-distribution detection on graphs. In NeurIPS, pages 30277–30290,

    2022

  33. [2020]

    End-to-end open-set semi-supervised node classification with out-of-distribution detection

    [Huang et al., 2022] Tiancheng Huang, Donglin Wang, and Yuan Fang. End-to-end open-set semi-supervised node classification with out-of-distribution detection. In IJCAI,

    2022

  34. [2021]

    Twitch gamers: A dataset for evaluating proximity preserving and structural role-based node embeddings

    [Rozemberczki and Sarkar, 2021] Benedek Rozemberczki and Rik Sarkar. Twitch gamers: A dataset for evaluating proximity preserving and structural role-based node embeddings. arXiv preprint arXiv:2101.03091,

    2021 arXiv

  35. [2022]

    A data-centric framework to endow graph neural networks with out-of- distribution detection ability

    [Guo et al., 2023] Yuxin Guo, Cheng Yang, Yuluo Chen, Jixi Liu, Chuan Shi, and Junping Du. A data-centric framework to endow graph neural networks with out-of- distribution detection ability. In KDD, pages 638–648,

    2023

  36. [2023]

    Deep gaussian embedding of graphs: Unsupervised inductive learning via ranking

    [Bojchevski and G¨unnemann, 2017] Aleksandar Bojchevski and Stephan G ¨unnemann. Deep gaussian embedding of graphs: Unsupervised inductive learning via ranking. arXiv preprint arXiv:1707.03815,

    2017 arXiv

  37. [2024]

    Evaluating robustness and uncer- tainty of graph models under structural distributional shifts

    [Bazhenov et al., 2023] Gleb Bazhenov, Denis Kuznedelev, Andrey Malinin, Artem Babenko, and Liudmila Prokhorenkova. Evaluating robustness and uncer- tainty of graph models under structural distributional shifts. NeurIPS, pages 75567–75594,

    2023

  38. [2025]

    Decoupled graph energy-based model for node out-of-distribution de- tection on heterophilic graphs

    [Chen et al., 2025] Yuhan Chen, Yihong Luo, Yifan Song, Pengwen Dai, Jing Tang, and Xiaochun Cao. Decoupled graph energy-based model for node out-of-distribution de- tection on heterophilic graphs. In ICLR,

    2025

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.