Pith. sign in

REVIEW 4 major objections 5 minor 43 references

EVINET: Towards Open-World Graph Learning via Evidential Reasoning Network

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A single graph model flags wrong predictions and novel classes through evidential reasoning.

desk verdict Solid empirical paper on joint misclassification and OOD detection for graphs; the logical-reasoning story is not backed by evidence, but the method works and the experiments are careful enough to merit a serious review. read the letter →

arxiv 2506.07288 v3 pith:WZAA663R submitted 2025-06-08 cs.LG

Weijie Guan , Haohui Wang , Jian Kang , Lihui Liu , Dawei Zhou This is my paper
classification cs.LG
keywords open-worldgraphlearningout-of-distributiondetectionmisclassificationevidentialreasoningBetaembeddingsubjectivelogicuncertaintyestimationneuralnetworks
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

EviNet sets out to solve two failure modes that appear when graph-learning models leave the closed world of fixed labels: wrong predictions on in-distribution nodes, and encounters with nodes from classes never seen in training. The paper argues that both can be handled by one framework if the model represents nodes as Beta distributions, reasons about classes with logical disjunction and negation, and converts the resulting evidence into subjective-logic uncertainty scores. It claims that a dissonance score detects misclassified nodes and a vacuity score detects out-of-distribution nodes, and that the framework keeps in-distribution accuracy competitive while improving both detection tasks. The practical stake is that a model deployed in an open setting could tell its user 'this prediction is unreliable' and 'this node does not fit any known class' instead of silently failing on novel fraud patterns or noisy transactions.

What carries the argument

The load-bearing object is the Beta embedding, which represents each node as a collection of independent Beta distributions instead of a point vector and is closed under the logical operations EviNet needs. Class embeddings are formed by a neural disjunction of the Beta embeddings of training nodes, and the set of all known classes is negated parameter-wise to build an implicit 'novel class' support region; a learnable per-node prior weight is produced by a class-specific GCN that reads the node together with that novel-class embedding. Subjective logic then converts per-class evidence into belief, uncertainty, dissonance, and vacuity scores. The trainable prior weight replaces the fixed prior of earlier subjective-logic methods, which the paper argues causes overconfidence, and a KL-based phase trains the node and class embeddings while an expected cross-entropy phase trains the evidence estimator.

What would settle it

Take a fixed graph dataset and swap only the held-out OOD classes, keeping the in-distribution classes identical; if vacuity-based OOD AUROC drops sharply or the ranking of OOD nodes becomes random, the novel-class support region is not capturing novelty itself but something about the specific held-out classes. A cheaper check is to measure vacuity on in-distribution nodes that sit near decision boundaries: if their vacuity is as high as genuine OOD nodes, the score is tracking geometric marginality rather than class novelty.

Watch

Extended reading notes

Core claim

EviNet's central claim is that explicit logical reasoning over Beta embeddings, mediated by subjective logic, gives graph models reliable uncertainty signals for both misclassification detection and out-of-distribution detection. Nodes are encoded as collections of Beta distributions; each known class is represented by the logical disjunction of the embeddings of its training nodes; and the region of all known classes is negated to produce an implicit support region for novel classes. A per-node, context-aware prior weight is learned from that novel-class region, the dissonance score flags nodes with conflicting evidence across known classes, and the vacuity score flags nodes whose evidence is concentrated in the novel region. Across five benchmark graphs the paper reports the lowest misclassification AURC on all five, the best OOD AUROC on four of five, and in-distribution accuracy competitive with the strongest baselines. The broader claim is that uncertainty estimation plus logical reasoning, rather than more model capacity, is what drives reliable open-world behavior.

Load-bearing premise

The paper assumes that the implicit novel-class region, formed simply by negating the disjunction of known-class embeddings, really does stand in for unseen classes, so the learned prior weight will give out-of-distribution nodes high vacuity even though the model never trains on such nodes.

Editorial extensions

If this is right

  • A deployed model could return a misclassification alert and an out-of-distribution alert from the same forward pass, so operators do not need separate detectors for the two failure modes.
  • Because the learnable prior weight is trained only on in-distribution data yet the method reports strong OOD detection, the vacuity score would serve as a practical novelty alarm without access to novel-class examples during training.
  • Replacing the fixed prior weight with a per-node learned one would remove the main overconfidence failure the paper attributes to prior subjective-logic graph models.
  • The consistent AURC gains across all five datasets indicate the dissonance signal transfers across different graph domains and label structures.
  • The approach keeps in-distribution accuracy competitive, so adopting uncertainty-aware reasoning need not come at the cost of ordinary classification performance.

Reading between the lines

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

  • I read the novel-class support region as a geometric proxy rather than a true semantic description of unseen classes; the method's OOD power may come mostly from the learned prior weight adapting vacuity to the observed decision boundary.
  • The same disjunction-and-negation machinery could be applied to other probabilistic embeddings, such as Gaussian or Von Mises-Fisher embeddings, which would test whether Beta closure is essential or whether any distribution family with logical operators works.
  • A natural extension is to make the novel-class region depend on local neighborhood context rather than global known-class disjunction, which could help in graphs where OOD nodes cluster in specific regions.
  • If the vacuity signal is genuinely driven by the geometric complement of known classes, then fine-grained or semantically close novel classes should be the hardest case; that is a testable prediction the paper does not run.
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

4 major / 5 minor

Summary. The paper proposes EviNet, a graph learning framework for open-world settings that addresses in-distribution misclassification detection and out-of-distribution (OOD) detection in a single model. Nodes and classes are represented as Beta embeddings, and subjective-logic-inspired uncertainty scores are computed: a dissonance score for misclassification detection and a vacuity score for OOD detection. The vacuity score uses a learnable per-node prior weight W_i that is derived from a class embedding for novel classes, obtained by applying logical disjunction and negation to the known-class embeddings. The authors report experiments on five graph datasets against nine baselines, showing that EviNet achieves the lowest AURC on all datasets and the best OOD AUROC on four of five datasets, alongside ablations, hyperparameter sensitivity, and scalability experiments.

Significance. If the central claim holds, EviNet would be a practical contribution: a single framework that provides calibrated uncertainty for both misclassification and OOD detection on graphs, with empirical gains over existing methods. The paper is strong on experimental grounding: five datasets, five runs with standard deviations, ablations separating the main modules, hyperparameter analysis for the only novel hyperparameter, and a scalability study on ogbn-arxiv. The code is available. However, the distinctive theoretical claim—that an implicit logical support region for novel classes, constructed by negating known-class Beta embeddings, is what drives OOD detection—is not adequately validated, because the negation operator is not derived, the definition of the novel-class region is garbled, and the vacuity prior is trained only with in-distribution supervision.

major comments (4)
  1. [Sec. 3.3, Eq. (10)] Equation (10) defines the novel-class support region as C_Nov = ∪_{k=1}^K C_k, which is exactly the same expression as C_Known in Eq. (9); the negation operation is never applied in the displayed formula. Since C_Nov is the foundation of the Vacuity Reasoning module, this is a load-bearing definitional error, not a typographical nit. Please correct the equation to C_Nov = NEGATION(C_Known) (or an equivalent explicit form) and ensure all downstream uses, including the construction of X_Nov in Eq. (12), are consistent with the corrected definition.
  2. [Sec. 3.3, Eq. (11)] The negation operator defined as NEGATION(α_i, β_i) = (1/α_i, 1/β_i) is not the standard Beta-embedding negation (β_i, α_i) from the cited BetaE framework [24], and no probabilistic or logical derivation is provided. For a Beta distribution, the natural complement is given by (β, α) because if X ~ Beta(α, β) then 1-X ~ Beta(β, α); the reciprocal parameterization has no such interpretation. As the entire novel-class support region rests on this operator, the paper must either derive the reciprocal form from a principled notion of Beta-region complement, or explicitly state that it is a new heuristic and provide evidence that it behaves as a complement.
  3. [Sec. 3.3-3.4, Eqs. (12), (15), (17)] The learnable prior weight W_i, computed by GCN_Nov from the novel-class embedding C_Nov, is trained exclusively by the in-distribution cross-entropy loss of Eq. (17); no loss term supervises W_i with OOD examples, and no regularization ties W_i to the geometric content of C_Nov. Consequently, the substantial OOD improvements in the ablation rows (c)-(e) of Table 2 could be produced by the extra GCN_Nov parameters learning a generic OOD-ranking confidence score that ignores the logical construction entirely. To support the paper's central claim that logical reasoning about novel classes is what drives OOD detection, please add a controlled experiment in which C_Nov is replaced by a random or fixed embedding (or the negation is removed) while keeping the same architecture and training objective; if performance does not drop, the logical construction is not empirically load-bearing.
  4. [Sec. 3.4, text after Eq. (15)] The text states 'By including the embeddings of other classes (including novel classes) as negative samples,' but the negative-sum term in Eq. (15) ranges only over known classes k ≠ y_i, and C_Nov is not available during phase-one training because it depends on the class embeddings C_k that are themselves learned. This is an internal inconsistency in the description of the training objective. Please revise the sentence to say 'other known classes' or clarify how novel-class embeddings participate in phase one.
minor comments (5)
  1. [Sec. 3.3, Eq. (10)] Even after correcting the missing negation, the notation around Eq. (9)-Eq. (11) is confusing: Eq. (10) currently says C_Nov = C_Known, and the negation is only mentioned in prose. Please restate the definitions so that C_Known, the negation operator, and C_Nov are each given distinct, unambiguous formulas.
  2. [Sec. 4.1, Dataset creation] The label leave-out strategy for creating OOD sets is described only briefly. Please specify which classes are held out for each dataset, whether the OOD nodes are present in the graph adjacency structure during training (transductive leakage), and whether any OOD labels are used for model selection.
  3. [Sec. 4.1, Implementation Details] The model selection criterion is given as 'Acc + AUROC − 10*AURC' but it is unclear whether this same criterion is applied to all baselines or only to EviNet. Please clarify, because differences in model selection could affect the comparisons in Table 1.
  4. [Sec. 3.2, Eq. (5)] The disjunction operator in Eq. (5) applies Softplus to a projected and weighted average, but it is not stated whether Softplus is applied elementwise to a vector and how the resulting parameters are clamped to be positive. A brief note would improve reproducibility.
  5. [Appendix D, Table 6] The misclassification AUROC for EviNet (78.96 ± 1.88) is substantially lower than for CRL (87.04 ± 2.15) and GKDE (87.33 ± 2.75) under validation-accuracy selection. The paper dismisses AUROC for misclassification detection as misleading, but readers may still find this gap surprising; please add a sentence explaining why AURC is the appropriate primary metric in this comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EviNet's uncertainty scores are learned model outputs evaluated against held-out OOD data, not quantities that reduce by construction to their training inputs.

full rationale

EviNet's derivation chain does not contain a step in which a predicted quantity is defined in terms of the same quantity, nor a fitted parameter renamed as a prediction. The class embeddings C_k are disjunctions of the same node embeddings being trained (Eq. 6), and the novel-class support region C_Nov is specified as the negation of the known-class union (Eqs. 9-11, though Eq. 10 as printed omits the NEGATION symbol), but these are architectural definitions used to construct features, not outputs whose evaluation is forced by the definitions; the reported OOD and misclassification metrics are measured on held-out nodes and held-out OOD classes under a label leave-out protocol (Section 4.1). The prior weight W_i is a trainable output of GCN_Nov (Eq. 12) and is supervised only through the in-distribution classification loss (Eq. 17), so the vacuity score is learned without OOD labels; this makes the 'logical reasoning' interpretation an empirical claim rather than a circular one. The only arguably nonstandard element, the reciprocal negation (1/alpha, 1/beta) in Eq. (11), is presented as the paper's own definition and is not imported from an external uniqueness theorem, so it is a correctness or semantic concern rather than circularity. Self-citations (e.g., refs. 30-32) appear in related work and are not load-bearing premises in the derivation. Appendix D's caveat that AUROC and AUPR 'can be misleading for misclassification detection' similarly concerns metric choice, not circularity. Accordingly no circular step is found.

Assumptions & free parameters 1 free parameters · 4 assumptions · 1 invented entities

The central claim rests on the learned prior weight and the negation-based novel class embedding, both introduced ad hoc. The paper provides no derivation that the vacuity score is calibrated for OOD nodes, and the model selection criterion may involve OOD validation labels that are not described.

free parameters (1)
  • gamma (margin in BL loss) = 55, 15, 55, 15, 15 for Amazon-photo, Amazon-computer, Coauthor-cs, Coauthor-physics, Wiki-cs respectively (Table 7)
    Margin in the Beta embedding ranking loss (Eq. 15); selected by grid search over {15, 55, 95, 135} on the validation set.
assumptions (4)
  • ad hoc to paper Beta distributions can represent node and class embeddings such that logical disjunction and negation operations produce meaningful support regions for known and novel classes.
    Invoked in Section 3.2 Eq. (5)-(6) and Section 3.3 Eq. (9)-(11). The paper relies on Beta embedding but extends it to node classification with a neural disjunction and a parameter-wise reciprocal negation; no theoretical guarantee is provided that these operations preserve class geometry.
  • ad hoc to paper The KL divergence between Beta embeddings is an appropriate distance for aligning nodes with their class embedding and separating from other classes.
    Used in the BL loss Eq. (15)-(16). This is a modeling choice from the Beta embedding literature, not derived for the graph node setting.
  • domain assumption The graph is an undirected attributed graph, and transductive node classification with label leave-out is a valid proxy for open-world graph learning.
    Stated in Problem 1, Section 2, and experiments Section 4.1. The open-world claim is tested by holding out entire classes, which assumes the OOD distribution is class-structured rather than covariate-shift-like.
  • standard math Subjective logic equations (1)-(4), including the definitions of vacuity and dissonance, are accepted as standard.
    Preliminary Section 2, Eq. (1)-(4). Standard subjective logic from the cited reference [14].
invented entities (1)
  • Implicit support region for novel classes (CNov)
    purpose: Serves as a representation of out-of-distribution classes, computed as the negation of the disjunction of known-class embeddings, and used as input to GCN_Nov for computing the prior weight W_i and vacuity score.
    Defined in Eq. (10)-(11). No external evidence shows this reciprocal Beta region corresponds to actual novel classes; the only support is the paper's internal ablation (Table 2), which shows OOD metrics improve when M2 is included.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EVINET: Towards Open-World Graph Learning via Evidential Reasoning Network." pith.science (2026). https://pith.science/paper/WZAA663R

@misc{pith2026250607288,
  author       = {Pith},
  title        = {Pith review of: EVINET: Towards Open-World Graph Learning via Evidential Reasoning Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WZAA663R}},
  note         = {Machine review of arXiv:2506.07288}
}
read the original abstract

Graph learning has been crucial to many real-world tasks, but they are often studied with a closed-world assumption, with all possible labels of data known a priori. To enable effective graph learning in an open and noisy environment, it is critical to inform the model users when the model makes a wrong prediction to in-distribution data of a known class, i.e., misclassification detection or when the model encounters out-of-distribution from novel classes, i.e., out-of-distribution detection. This paper introduces Evidential Reasoning Network (EVINET), a framework that addresses these two challenges by integrating Beta embedding within a subjective logic framework. EVINET includes two key modules: Dissonance Reasoning for misclassification detection and Vacuity Reasoning for out-of-distribution detection. Extensive experiments demonstrate that EVINET outperforms state-of-the-art methods across multiple metrics in the tasks of in-distribution classification, misclassification detection, and out-of-distribution detection. EVINET demonstrates the necessity of uncertainty estimation and logical reasoning for misclassification detection and out-of-distribution detection and paves the way for open-world graph learning. Our code and data are available at https://github.com/SSSKJ/EviNET.

Figures

Figures reproduced from arXiv: 2506.07288 by the authors.

Figure 1
Figure 1. An illustrative figure of financial fraud situations [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of EviNet. The framework processes graph-structured data through two core modules: (1) Dissonance Reasoning (M1); (2) Vacuity Reasoning (M2). The final outputs include in-distribution classification, misclassification detection scores, and OOD detection scores. For simplicity, we consider a binary classification scenario in this example, in which Class 1 is shown in blue, Class 2 is shown in green, and N… view at source ↗
Figure 3
Figure 3. An illustrative figure for M1. Dissonance Reasoning [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Hyperparameter analysis for 𝛾 on Amazon-photo accuracy (Acc), AURC, FPR95, and AUROC across different values of 𝛾, demonstrating that EviNet is generally robust to 𝛾. Remark: We also present the analysis of the relationship between the running time and the number of no…
Figure 5
Figure 5. Figure 5: Running time vs. the number of nodes and edge [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 28 canonical work pages

  1. [24]

    Hongyu Ren and Jure Leskovec. 2020. Beta Embeddings for Multi-Hop Logical Reasoning in Knowledge Graphs. In Advances in Neural Informa- tion Processing Systems 33: Annual Conference on Neural Information Pro- cessing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual , Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hs...

    work page 2020

  2. [1]

    Ibrahem Al-Obaydy and Shahrel Azmin Suandi

    Wasseem N. Ibrahem Al-Obaydy and Shahrel Azmin Suandi. 2020. Open-Set Face Recognition in Video Surveillance: A Survey . 425–436. doi:10.1007/978-981-15- 2317-5_36

    doi:10.1007/978-981-15- 2020

  3. [2]

    2015.Support Vector Machines for Classification

    Mariette Awad and Rahul Khanna. 2015.Support Vector Machines for Classification. Apress, Berkeley, CA, 39–66. doi:10.1007/978-1-4302-5990-9_3

    work page doi:10.1007/978-1-4302-5990-9_3 2015

  4. [3]

    Bastian, and Gokhan Kul

    Gaspard Baye, Priscila Silva, Alexandre Broggi, Lance Fiondella, Nathaniel D. Bastian, and Gokhan Kul. 2023. Performance Analysis of Deep-Learning Based Open Set Recognition Algorithms for Network Intrusion Detection Systems. In NOMS 2023-2023 IEEE/IFIP Network Operations and Management Symposium . 1–6. doi:10.1109/NOMS56928.2023.10154410

    arXiv 2023

  5. [4]

    Alexander Cao, Diego Klabjan, and Yuan Luo. 2023. Open-set recognition of breast cancer treatments. Artificial Intelligence in Medicine 135 (2023), 102451. doi:10.1016/j.artmed.2022.102451

    work page arXiv 2023

  6. [5]

    Mengyuan Chen, Junyu Gao, and Changsheng Xu. 2024. R-EDL: Relaxing Nonessential Settings of Evidential Deep Learning. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11,

    work page 2024

  7. [6]

    P ERDdS and A R&wi. 1959. On random graphs I. Publ. math. debrecen 6, 290-297 (1959), 18

    work page 1959

  8. [7]

    Yonatan Geifman and Ran El-Yaniv. 2017. Selective Classification for Deep Neural Networks. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA , Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, ...

    work page 2017

Show all 43 references
  1. [8]

    Federica Granese, Marco Romanelli, Daniele Gorla, Catuscia Palamidessi, and Pablo Piantanida. 2021. DOCTOR: A Simple Method for Detecting Misclassifica- tion Errors. In Advances in Neural Information Processing Systems , M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and ...

    2021

  2. [9]

    Aditya Grover and Jure Leskovec. 2016. node2vec: Scalable Feature Learning for Networks. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016, Balaji Krishnapuram, Mohak Shah, Alexande...

    2016

  3. [10]

    Rudd, and Terrance E

    Manuel Günther, Steve Cruz, Ethan M. Rudd, and Terrance E. Boult. 2017. Toward Open-Set Face Recognition. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2017, Honolulu, HI, USA, July 21-26, 2017. IEEE Computer Society, 573–582. doi...

    2017 doi

  4. [11]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive Representation Learning on Large Graphs. In Advances in Neural Information Processing Systems , I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (Eds.), Vol. 30. Curran As...

    2017

  5. [12]

    Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joseph Kwon, Mo- hammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. 2022. Scaling Out-of- Distribution Detection for Real-World Settings. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Ba...

    2022

  6. [13]

    Hitham Jleed and Martin Bouchard. 2020. Open Set Audio Recognition for Multi-Class Classification With Rejection. IEEE Access 8 (2020), 146523–146534. doi:10.1109/ACCESS.2020.3015227

    2020

  7. [14]

    Audun Josang, Jin-Hee Cho, and Feng Chen. 2018. Uncertainty characteristics of subjective opinions. In 2018 21st International Conference on Information Fusion (FUSION). IEEE, 1998–2005

    2018

  8. [15]

    Kazi Zainab Khanam, Gautam Srivastava, and Vijay Mago. 2023. The homophily principle in social network analysis: A survey. Multimedia Tools and Applications 82, 6 (2023), 8811–8854

    2023

  9. [16]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2016. Variational Graph Auto-Encoders. arXiv:1611.07308 [stat.ML] https://arxiv.org/abs/1611.07308

    2016 arXiv

  10. [17]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net. https://openreview.n...

    2017

  11. [18]

    Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. 2018. A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adver- sarial Attacks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, N...

    2018

  12. [19]

    Xue Li, Jinlong Fei, Jiangtao Xie, Ding Li, Heng Jiang, Ruonan Wang, and Zan Qi. 2023. Open Set Recognition for Malware Traffic via Predictive Uncertainty. Electronics 12, 2 (2023). https://www.mdpi.com/2079-9292/12/2/323

    2023

  13. [20]

    Shiyu Liang, Yixuan Li, and R. Srikant. 2018. Enhancing The Reliability of Out-of- distribution Image Detection in Neural Networks. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedin...

    2018

  14. [21]

    Owens, and Yixuan Li

    Weitang Liu, Xiaoyun Wang, John D. Owens, and Yixuan Li. 2020. Energy-based Out-of-distribution Detection. In Advances in Neural Infor- mation Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual , H...

    2020

  15. [22]

    Péter Mernyei and Cătălina Cangea. 2022. Wiki-CS: A Wikipedia-Based Bench- mark for Graph Neural Networks. arXiv:2007.02901 [cs.LG] https://arxiv.org/ abs/2007.02901

    2022 arXiv

  16. [23]

    Jooyoung Moon, Jihyo Kim, Younghak Shin, and Sangheum Hwang. 2020. Confidence-Aware Learning for Deep Neural Networks. InProceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event (Proceedings of Machine Learning Research, V...

    2020

  17. [25]

    Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. 2019. Pitfalls of Graph Neural Network Evaluation. arXiv:1811.05868 [cs.LG] https://arxiv.org/abs/1811.05868

    2019 arXiv

  18. [26]

    Gun-Yoon Shin, Dong-Wook Kim, and Myung-Mook Han. 2023. Open Set Recog- nition With Dissimilarity Weight for Unknown Attack Detection. IEEE Access 11 (2023), 102381–102390. doi:10.1109/ACCESS.2023.3277871

    2023

  19. [27]

    Maximilian Stadler, Bertrand Charpentier, Simon Geisler, Daniel Zügner, and Stephan Günnemann. 2021. Graph Posterior Network: Bayesian Predictive Uncertainty for Node Classification. In Advances in Neural Information Pro- cessing Systems 34: Annual Conference on Neural Informa...

    2021

  20. [28]

    Artem Vazhentsev, Gleb Kuzmin, Artem Shelmanov, Akim Tsvigun, Evgenii Tsymbalov, Kirill Fedyanin, Maxim Panov, Alexander Panchenko, Gleb Gusev, Mikhail Burtsev, Manvel Avetisian, and Leonid Zhukov. 2022. Uncertainty Esti- mation of Transformer Predictions for Misclassification...

    2022

  21. [29]

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph Attention Networks. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedi...

    2018

  22. [30]

    Haohui Wang, Weijie Guan, Jianpeng Chen, Zi Wang, and Dawei Zhou

  23. [31]

    Haohui Wang, Baoyu Jing, Kaize Ding, Yada Zhu, Wei Cheng, Si Zhang, Yonghui Fan, Liqing Zhang, and Dawei Zhou. 2024. Mastering Long-Tail Complexity on Graphs: Characterization, Learning, and Generalization. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discover...

    2024

  24. [32]

    Towards Heterogeneous Long-tailed Learning: Benchmarking, Metrics, and Toolbox. In Advances in Neural Information Processing Sys- tems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024 , Amir Glober...

    2024

  25. [33]

    Jianian Wang, Sheng Zhang, Yanghua Xiao, and Rui Song. 2022. A Review on Graph Neural Network Methods in Financial Applications. arXiv:2111.15367 [q- fin.ST]

    2022 arXiv

  26. [34]

    Haohui Wang, Yuzhen Mao, Yujun Yan, Yaoqing Yang, Jianhui Sun, Kevin Choi, Balaji Veeramani, Alison Hu, Edward Bowen, Tyler Cody, and Dawei Zhou

  27. [35]

    In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024

    EvoluNet: Advancing Dynamic Non-IID Transfer Learning on Graphs. In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024 . OpenReview.net. https://openreview.net/forum?id= anM1M5aoM8

    2024

  28. [36]

    Qitian Wu, Yiting Chen, Chenxiao Yang, and Junchi Yan. 2023. Energy-based Out-of-Distribution Detection for Graph Neural Networks. In The Eleventh Inter- national Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net. https://openrevi...

    2023

  29. [37]

    Tuo Wang, Jian Kang, Yujun Yan, Adithya Kulkarni, and Dawei Zhou. 2025. Non-exchangeable Conformal Prediction for Temporal Graph Neural Networks. In Proceedings of the 31st ACM SIGKDD International conference on knowledge discovery & data mining

    2025

  30. [38]

    Longfeng Wu, Yao Zhou, Jian Kang, and Dawei Zhou. 2025. Bridging Fairness and Uncertainty: Theoretical Insights and Practical Strategies for Equalized Coverage in GNNs. In THE WEB CONFERENCE 2025 . https://openreview.net/forum?id= lZQp7s69ls

    2025

  31. [39]

    Chen Zhao, Dawei Du, Anthony Hoogs, and Christopher Funk. 2023. Open Set Action Recognition via Multi-Label Evidential Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . 22982–22991

    2023

  32. [40]

    Shiwen Wu, Fei Sun, Wentao Zhang, Xu Xie, and Bin Cui. 2022. Graph Neural Networks in Recommender Systems: A Survey. ACM Comput. Surv. 55, 5, Article 97 (dec 2022), 37 pages. doi:10.1145/3535101

    2022 doi

  33. [41]

    Jiawen Xu, Matthias Kovatsch, and Sergio Lucia. 2021. Open Set Recognition for Machinery Fault Diagnosis. In 2021 IEEE 19th International Conference on Industrial Informatics (INDIN). 1–7. doi:10.1109/INDIN45523.2021.9557572

    2021

  34. [43]

    Xujiang Zhao, Feng Chen, Shu Hu, and Jin-Hee Cho. 2020. Uncertainty Aware Semi-Supervised Learning on Graph Data. In Advances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (Eds.), Vol. 33. Curran Associates, Inc., 1282...

    2020

  35. [2024]

    https://openreview.net/forum?id=Si3YFA641c

    OpenReview.net. https://openreview.net/forum?id=Si3YFA641c

Pith tools

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