Pith. sign in

REVIEW 3 major objections 4 minor 73 references

Uncertainty-Aware Graph Neural Networks: A Multi-Hop Evidence Fusion Approach

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read EFGNN claims that fusing evidence and uncertainty across multiple propagation depths yields more accurate and more trustworthy node classifications than any single depth, using a parameter-free cumulative belief fusion operator.

desk verdict EFGNN's accuracy gains are plausible and the CBF fusion is a clean parameter-free idea, but the trustworthiness framing rests on an uncertainty signal that is never calibrated; conditional accept, not reject. read the letter →

arxiv 2506.13083 v1 pith:WOL2PTIY submitted 2025-06-16 cs.LG

classification cs.LG
keywords graphneuralnetworksuncertaintyquantificationsubjectivelogicevidencefusionnodeclassificationmulti-hoppropagationtrustworthypredictioncumulativebelief
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 sets out to show that a graph neural network should not commit to a single propagation depth: predictions made from 1-hop, 2-hop, and deeper neighborhoods carry different and partly independent uncertainty signals, and the depth that gives the most confident prediction varies by node class. EFGNN therefore generates an evidence vector, a Dirichlet distribution, and a subjective-logic opinion at every propagation step, then merges all of them with a parameter-free cumulative belief fusion (CBF) operator. The authors claim this fusion yields both higher node-classification accuracy and a usable uncertainty score for each final prediction. They report top accuracy on Cora, Pubmed, Computers, and Photo, and argue that the theoretical propositions on CBF guarantee that fusing cannot increase uncertainty and can raise belief in the true class under a stated condition.

What carries the argument

The load-bearing object is the cumulative belief fusion (CBF) operator: a parameter-free rule for merging the subjective-logic opinions $\omega^{\ell}=(b^{\ell},u^{\ell},a^{\ell})$ produced at different propagation depths. It is implemented as simple addition of evidence across hops, $\hat e_{ik}=\sum_{\ell}e^{\ell}_{ik}$, from which the joint belief and uncertainty masses follow; when two opinions are fused directly, the operator takes the explicit closed form in Eq. (16) with $\hat u = u^{\xi}u^{\varphi}/(u^{\xi}+u^{\varphi}-u^{\xi}u^{\varphi})$ and a corresponding belief update. This operator carries the entire argument: it is what lets the model keep every hop's evidence instead of selecting one depth, and the paper's theoretical propositions (fusion never increases uncertainty; fusion can increase true-class belief; harm is bounded when one view is certain) all follow from its algebraic form. It is also what makes the method cheap, since the fusion cost is $O(nK)$ regardless of the number of hops.

What would settle it

Compute a reliability diagram and expected calibration error for EFGNN's fused uncertainty on held-out nodes. The central claim is falsified if low-uncertainty predictions are not systematically more accurate than high-uncertainty ones, or if per-node checks show many nodes where the Proposition 3 condition $b_t^{\xi} \ge b_m^{\varphi}$ fails yet CBF still appears to help, since that would mean the reported gains come from ensemble averaging rather than from the evidence-fusion mechanism.

Watch

Extended reading notes

Core claim

The paper's central claim is that prediction uncertainty in GNNs is depth-dependent and that this dependence can be exploited rather than ignored. On eight benchmarks, the proposed EFGNN learns a shared evidence network that maps the propagated features at each hop to nonnegative evidence $e^{\ell}$, forms the Dirichlet parameter $\alpha^{\ell}=e^{\ell}+1$, and reads off a belief mass $b^{\ell}_{k}=e^{\ell}_{k}/S^{\ell}$ and an uncertainty mass $u^{\ell}=K/S^{\ell}$ for each hop. The cumulative belief fusion operator then combines the per-hop opinions into one joint opinion: the joint evidence is the simple sum of per-hop evidence, giving fused uncertainty $\hat u = u^{\xi}u^{\varphi}/(u^{\xi}+u^{\varphi}-u^{\xi}u^{\varphi})$. The paper proves that this fused uncertainty is never larger than either input uncertainty and that, under the condition that the best single-depth belief in the true class does not exceed the other view's true-class belief, fusion raises the fused true-class belief; it also notes that performance loss is bounded when one view is already fully certain. Empirically, EFGNN reports the highest test accuracy on four of eight datasets and remains accurate at 16, 32, and 64 propagation steps where plain GCN degrades sharply.

Load-bearing premise

The trustworthiness claim depends on the uncertainty score $u=K/S$ being a genuinely faithful measure of prediction risk; the paper validates this mainly with threshold-accuracy curves and a qualitative out-of-distribution density plot, and reports no calibration metric, so if the score is miscalibrated on real deployment distributions the risk-flagging claim fails even if the accuracy tables hold.

Editorial extensions

If this is right

  • The model keeps high accuracy at depths where plain GCN collapses, because no single hop must be the right one: CBF accumulates evidence from all of them.
  • A user of EFGNN can threshold the fused uncertainty score and keep only predictions below the threshold, with accuracy of the selected set rising as the threshold is tightened.
  • Because fused uncertainty is provably no larger than any input hop's uncertainty, adding more propagation views cannot make the model less certain even when the extra views are low quality.
  • The fusion step needs no trained parameters and costs $O(nK)$, so the trustworthiness machinery adds almost no overhead on top of multi-hop propagation.
  • The wrong-prediction risk is made explicit: the model outputs a per-node uncertainty that the paper uses to flag risky predictions and to detect Gaussian-corrupted out-of-distribution inputs.

Reading between the lines

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

  • Beyond the paper: if the uncertainty scores are calibrated in deployment, the same architecture could be used for selective prediction or human-in-the-loop deferral, holding back nodes with $\hat u$ above a threshold and accepting the rest at a higher accuracy.
  • Beyond the paper: the CBF algebra does not care how the opinions were generated, so the same operator could fuse evidence from different graph perturbations, feature subsets, or even different source graphs, and Proposition 1 would still apply.
  • Beyond the paper: a sharper test of the mechanism would compare CBF against simple averaging of per-hop softmax probabilities; if averaged probabilities match CBF accuracy, the Dirichlet and subjective-logic machinery is not the source of the gain, and if CBF clearly wins, the uncertainty weighting is doing real work.
  • Beyond the paper: the depth-analysis result suggests a complementary design the paper does not explore, namely per-class or per-node depth selection before fusion, to concentrate evidence on the hops that already carry the strongest signal.
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 / 4 minor

Summary. The paper proposes EFGNN, an uncertainty-aware graph neural network that generates subjective-logic opinions at multiple propagation depths and fuses them with a cumulative belief fusion (CBF) operator. Per-hop evidence is produced by a shared softplus network on decoupled propagated features; CBF then sums evidence across hops, yielding a joint Dirichlet opinion whose uncertainty is claimed to quantify the risk of wrong predictions. Training combines an evidence cross-entropy loss, a dissonance coefficient regularizer, and a KL divergence regularizer. Experiments on eight graph datasets report accuracy gains over several GNN and uncertainty-aware baselines, with ablations showing that CBF contributes the largest improvement. Theoretical propositions state that fused uncertainty is no larger than per-hop uncertainty, that uncertainty is monotone in per-hop uncertainty, and that fusion can improve or not degrade accuracy under conditions on belief masses. The paper also reports robustness to depth and qualitative OOD uncertainty separation.

Significance. If the uncertainty estimate is well-calibrated, the paper offers a simple and practical way to combine multi-hop evidence with an explicit risk score, and the release of source code is a valuable asset. The accuracy results on several benchmarks are competitive, and the ablations provide clear evidence that the fusion mechanism is responsible for much of the gain. The CBF operator is parameter-free and computationally efficient, which is a genuine advantage. However, the headline claim of 'trustworthy prediction' and 'explicit risk of wrong predictions' is not yet supported: no calibration metrics are reported, and the threshold-accuracy analysis only demonstrates a ranking property, not calibration. In addition, one of the accuracy-theoretic propositions has an invalid proof step. The theoretical propositions are mostly elementary consequences of the fusion algebra and serve as internal consistency checks rather than deep new results.

major comments (3)
  1. [V-D, Eq. (20)] The proof of Proposition 4 contains an invalid inequality. In the second line of Eq. (20), the denominator is changed from (uξ+uφ−uξuφ) to (uξ+uφ). Since uξuφ≥0, we have uξ+uφ−uξuφ ≤ uξ+uφ, and therefore uξbφ^t/(uξ+uφ−uξuφ) ≥ uξbφ^t/(uξ+uφ). Consequently, bφ^t minus the former fraction is smaller than or equal to bφ^t minus the latter fraction, so the displayed chain of inequalities does not yield the claimed upper bound. The special case uφ=0 is correct, but the general bound 'performance degradation is limited' is not established by the argument as written. Please correct the proof or restate the proposition with a valid derivation.
  2. [VI-E, Fig. 10] The central trustworthiness claim is unsupported because no calibration evidence is provided. The paper defines fused uncertainty as u=K/S (Eq. 10) and uses it to argue that the model 'makes explicit the risk of wrong predictions.' Figures 10(a)–(e) show that lower uncertainty thresholds select higher-accuracy samples, which is a ranking-style check, but do not show that the uncertainty values are calibrated estimates of error probability. Since Proposition 1 guarantees that fused uncertainty is always smaller than each per-hop uncertainty, observing low fused uncertainty is a mechanical consequence of evidence accumulation and does not by itself indicate correctness. Please report expected calibration error (ECE) or reliability diagrams, and ideally a quantitative risk-coverage curve compared with baseline uncertainty methods, to substantiate the trustworthiness claim.
  3. [V-D, Proposition 3] Proposition 3 is stated under the condition bξ_t ≥ bφ_m, where t is the ground-truth class and bφ_m is the largest belief mass in view φ. This condition involves the unknown ground-truth label and is not verified empirically. The paper does not analyze what fraction of nodes in the trained models satisfy this condition, nor does it provide an argument that the training losses (Eq. 15) enforce it. Without such verification, the theoretical claim that fusion 'can improve the accuracy of the model' remains conditional in a way that is not shown to hold in practice. The ablation in Table III shows that CBF helps on average, but that is not the same as validating the proposition's assumption. Please add an empirical analysis of the condition's satisfaction rate, or weaken the theoretical claim accordingly.
minor comments (4)
  1. [V-A, Eq. (6)] The random perturbation formula appears to have a scaling error. If b is a Bernoulli mask with keep probability (1−σ), then the expectation-preserving scaling factor should be 1/(1−σ), not (1−σ). As written, E[\tilde Xℓ] = (1−σ)^2 Xℓ, which contradicts the statement that the expectation is unchanged. Please clarify the definition of σ or correct the scaling, and align the notation with Algorithm 1, which uses Dropout(X_t, b).
  2. [V-C, Eq. (14)] The KL divergence expression has a formatting typo: the left side reads 'KL [Dir (pi | ˜αi) || Dir (pi | 1) |' with a stray vertical bar, and the Γ-term is missing a closing bracket. Please correct the equation for clarity.
  3. [VI-A, Baselines] The text states that ten state-of-the-art GNNs are compared, but twelve are listed in Section VI-A (GCN, GAT, JK-Net, ResGCN, APPNP, AP-GCN, SGC, SIGN, S2GC, GAMLP, AGNN, Flip-APPNP). Additionally, the uncertainty-aware baseline list mentions Drop-GAT and S-BGAT-T, while Table II also includes Drop-GCN and S-BGCN-T. Please reconcile the count and the table entries.
  4. [IV-A] The phrase 'we use the true class probability as the confidence measure' is ambiguous. It presumably means the predicted probability assigned to the true class (or the model's confidence in the ground-truth label), but this should be stated explicitly to avoid confusion with an oracle measure.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EFGNN's fusion rule and theoretical propositions are derived from explicitly stated definitions and validated against external benchmarks; self-citations are not load-bearing.

full rationale

EFGNN's derivation chain is self-contained. The multi-hop evidence module (Eqs. 5-7), the cumulative belief fusion operator (Eqs. 9-10 and 16), and the joint objective (Eq. 15) are all specified in the paper, and the reported accuracy (Table II) and ablations (Table III) are measured against external datasets and baselines. No fitted parameter is renamed as a prediction: the CBF combination is parameter-free evidence addition, and the accuracy results are empirical rather than constructed from the model's own uncertainty metric. Propositions 1-4 are algebraic consequences of the fusion rule (Eq. 16) together with the subjective-logic identity u = K/S (Eq. 4); deriving internal consistency properties from an operator's definition is normal mathematical analysis, not circularity, and the paper does not present these propositions as empirical validation. The evidential-fusion formalism is credited to external prior work ([15], [51], [62]), while the only self-citations ([29], [30], [44]) appear in related-work context and do not carry the central argument. The lack of calibration metrics for the uncertainty score u = K/S, and the fact that fused uncertainty shrinks with the number of hops regardless of correctness, are legitimate correctness and validity concerns, but they are not cases where a claimed result reduces to its own input by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or model-level entities; its new ingredients are a fusion operator and a loss combination, not a new ontological object. The free parameters are standard tuning hyperparameters. The main assumptions imported from prior literature are the subjective-logic convention, the normalized-propagation schema, and the class-uniform base rate.

free parameters (4)
  • lambda_Dis (dissonance loss weight) = searched over {0.1, 0.3, 0.5}
    The weight for the dissonance regularization term in Eq. (15), tuned per dataset on the validation set.
  • lambda_KL (KL loss weight) = searched over {0.01, 0.05, 0.1}
    The weight for the KL divergence penalty in Eq. (15), tuned per dataset on the validation set.
  • Propagation depth T = searched over {2, 4, 6, 8, 16, 32}
    The number of propagation steps, tuned per dataset; the paper highlights that the optimal depth varies across datasets.
  • Perturbation probability b = searched over {0.1, 0.3, 0.5, 0.7}
    Dropout perturbation applied after propagation (Eq. (6)), tuned per dataset.
assumptions (5)
  • standard math The standard message-passing graph model with normalized self-looped adjacency matrix A-hat and linear propagation is sufficient to generate good node representations.
    Adopted without scrutiny from prior literature (Section III-B, Eq. (1)); this is standard GCN machinery whose limitations for heterophilic graphs are known.
  • standard math Subjective-logic opinion mapping Dir(p|alpha), with alpha_k = e_k + 1 and W=K, properly represents prediction uncertainty.
    Imported from Jøsang's subjective logic and evidential deep learning; this is the agreed convention in this literature stream.
  • domain assumption Identical base rate a = 1/K for all classes.
    Assumed without justification in Sections III-C and V-B; it is conventional but imposes a prior that may not reflect real class frequencies.
  • domain assumption The propagated features X_l, after random node dropout, give unbiased and informative per-depth evidence.
    The random perturbation in Eq. (6) is asserted to reduce sensitivity without proof; the empirical results support it indirectly, but the inductive bias from the perturbation to evidence quality is unverified.
  • standard math The CBF fusion of evidence from different depths is associative and commutes, ensuring the joint opinion is well-defined regardless of depth ordering.
    The theoretical guarantees in Section V-D invoke the four propositions from [15] and treat the fusion as an associative operator, but the actual addition of evidence vectors is associative by construction, so this is sound.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Uncertainty-Aware Graph Neural Networks: A Multi-Hop Evidence Fusion Approach." pith.science (2026). https://pith.science/paper/WOL2PTIY

@misc{pith2026250613083,
  author       = {Pith},
  title        = {Pith review of: Uncertainty-Aware Graph Neural Networks: A Multi-Hop Evidence Fusion Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WOL2PTIY}},
  note         = {Machine review of arXiv:2506.13083}
}
read the original abstract

Graph neural networks (GNNs) excel in graph representation learning by integrating graph structure and node features. Existing GNNs, unfortunately, fail to account for the uncertainty of class probabilities that vary with the depth of the model, leading to unreliable and risky predictions in real-world scenarios. To bridge the gap, in this paper, we propose a novel Evidence Fusing Graph Neural Network (EFGNN for short) to achieve trustworthy prediction, enhance node classification accuracy, and make explicit the risk of wrong predictions. In particular, we integrate the evidence theory with multi-hop propagation-based GNN architecture to quantify the prediction uncertainty of each node with the consideration of multiple receptive fields. Moreover, a parameter-free cumulative belief fusion (CBF) mechanism is developed to leverage the changes in prediction uncertainty and fuse the evidence to improve the trustworthiness of the final prediction. To effectively optimize the EFGNN model, we carefully design a joint learning objective composed of evidence cross-entropy, dissonance coefficient, and false confident penalty. The experimental results on various datasets and theoretical analyses demonstrate the effectiveness of the proposed model in terms of accuracy and trustworthiness, as well as its robustness to potential attacks. The source code of EFGNN is available at https://github.com/Shiy-Li/EFGNN.

Figures

Figures reproduced from arXiv: 2506.13083 by the authors.

Figure 1
Figure 1. Node embeddings from different hops of neighborhoods may exhibit [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The typical example of Dirichlet distribution and multinomial opinion. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An example of different optimal class-level certainty at GNNs with [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: The distributions of confidence w.r.t. true class probability for 6 classes (A-F) of Citeseer dataset at different depths. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: The distributions of confidence w.r.t. standard deviation density of class probability at different depths on 5 datasets. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: At the highest level, EFGNN is composed of two [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 6
Figure 6. Figure 6: Illustration of proposed method. Predictions with associated uncertainties are generated at multiple depths and then fused. The overall model flow [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Heat map visualization of the effects of two parameters [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: The standard deviation of the class probability distribution of multi-layer (2, 16, or 32 layers) GCN and EFGNN. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: The accuracy and uncertainty of EFGNN in different propagation [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Validity assessment of uncertainty on 5 datasets. (a)-(e): uncertainty threshold and accuracy; (f)-(j): uncertainty density distribution. [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Efficiency comparison among various baseline methods on the [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

73 extracted references · 47 canonical work pages

  1. [1]

    Graphboot: Quantifying uncertainty in node feature learning on large networks,

    C. G. Akcora, Y . R. Gel, M. Kantarcioglu, V . Lyubchich, and B. Thurais- ingham, “Graphboot: Quantifying uncertainty in node feature learning on large networks,” IEEE Transactions on Knowledge and Data Engi- neering, vol. 33, no. 1, pp. 116–127, 2019

  2. [2]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Ben- gio, “Graph attention networks,” arXiv preprint arXiv:1710.10903, 2017

  3. [3]

    Graph self-supervised learning: A survey,

    Y . Liu, M. Jin, S. Pan, C. Zhou, Y . Zheng, F. Xia, and S. Y . Philip, “Graph self-supervised learning: A survey,” IEEE Transactions on Knowledge and Data Engineering , vol. 35, no. 6, pp. 5879–5900, 2022

  4. [4]

    Geometric deep learning: Grids, groups, graphs, geodesics, and gauges,

    M. M. Bronstein, J. Bruna, T. Cohen, and P. Veliˇckovi´c, “Geometric deep learning: Grids, groups, graphs, geodesics, and gauges,” arXiv preprint arXiv:2104.13478, 2021

  5. [5]

    Accurate prediction of protein structures and interactions using a three- track neural network,

    M. Baek, F. DiMaio, I. Anishchenko, J. Dauparas, S. Ovchinnikov, G. R. Lee, J. Wang, Q. Cong, L. N. Kinch, R. D. Schaeffer et al. , “Accurate prediction of protein structures and interactions using a three- track neural network,” Science, vol. 373, no. 6557, pp. 871–876, 2021

  6. [6]

    Siren: Sign-aware recommendation using graph neural networks,

    C. Seo, K.-J. Jeong, S. Lim, and W.-Y . Shin, “Siren: Sign-aware recommendation using graph neural networks,” IEEE Transactions on Neural Networks and Learning Systems , 2022

  7. [7]

    Haqjsk: Hierarchical-aligned quantum jensen-shannon kernels for graph classification,

    L. Bai, L. Cui, Y . Wang, M. Li, J. Li, S. Y . Philip, and E. R. Hancock, “Haqjsk: Hierarchical-aligned quantum jensen-shannon kernels for graph classification,” IEEE Transactions on Knowledge and Data Engineering, 2024

  8. [8]

    Concrete problems in ai safety,

    D. Amodei, C. Olah, J. Steinhardt, P. Christiano, J. Schulman, and D. Man ´e, “Concrete problems in ai safety,” arXiv preprint arXiv:1606.06565, 2016

Show all 73 references
  1. [9]

    Trustworthy graph neural networks: Aspects, methods and trends,

    H. Zhang, B. Wu, X. Yuan, S. Pan, H. Tong, and J. Pei, “Trustworthy graph neural networks: Aspects, methods and trends,” arXiv preprint arXiv:2205.07424, 2022

  2. [10]

    On calibration of modern neural networks,

    C. Guo, G. Pleiss, Y . Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” in International conference on machine learning. PMLR, 2017, pp. 1321–1330

  3. [11]

    Posterior network: Uncertainty estimation without ood samples via density-based pseudo- counts,

    B. Charpentier, D. Z ¨ugner, and S. G ¨unnemann, “Posterior network: Uncertainty estimation without ood samples via density-based pseudo- counts,” Advances in Neural Information Processing Systems , vol. 33, pp. 1356–1367, 2020

  4. [12]

    Predictive uncertainty estimation via prior networks,

    A. Malinin and M. Gales, “Predictive uncertainty estimation via prior networks,” Advances in neural information processing systems , vol. 31, 2018

  5. [13]

    Bayesian graph convo- lutional neural networks for semi-supervised classification,

    Y . Zhang, S. Pal, M. Coates, and D. Ustebay, “Bayesian graph convo- lutional neural networks for semi-supervised classification,” in Proceed- ings of the AAAI conference on artificial intelligence , vol. 33, no. 01, 2019, pp. 5829–5836

  6. [14]

    R. M. Neal, Bayesian learning for neural networks . Springer Science & Business Media, 2012, vol. 118

  7. [15]

    Trusted multi-view classi- fication with dynamic evidential fusion,

    Z. Han, C. Zhang, H. Fu, and J. T. Zhou, “Trusted multi-view classi- fication with dynamic evidential fusion,” IEEE transactions on pattern analysis and machine intelligence , vol. 45, no. 2, pp. 2551–2566, 2022

  8. [16]

    Evidential deep learning to quantify classification uncertainty,

    M. Sensoy, L. Kaplan, and M. Kandemir, “Evidential deep learning to quantify classification uncertainty,” Advances in neural information processing systems, vol. 31, 2018

  9. [17]

    Trusted multi-view classifi- cation,

    Z. Han, C. Zhang, H. Fu, and J. T. Zhou, “Trusted multi-view classifi- cation,” arXiv preprint arXiv:2102.02051 , 2021

  10. [18]

    Uncertainty aware semi- supervised learning on graph data,

    X. Zhao, F. Chen, S. Hu, and J.-H. Cho, “Uncertainty aware semi- supervised learning on graph data,” Advances in Neural Information Processing Systems, vol. 33, pp. 12 827–12 836, 2020

  11. [19]

    Be confident! towards trust- worthy graph neural networks via confidence calibration,

    X. Wang, H. Liu, C. Shi, and C. Yang, “Be confident! towards trust- worthy graph neural networks via confidence calibration,” Advances in Neural Information Processing Systems , vol. 34, pp. 23 768–23 779, 2021

  12. [20]

    Uncertainty aware graph gaussian process for semi-supervised learning,

    Z.-Y . Liu, S.-Y . Li, S. Chen, Y . Hu, and S.-J. Huang, “Uncertainty aware graph gaussian process for semi-supervised learning,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 34, no. 04, 2020, pp. 4957–4964

  13. [21]

    Graph posterior network: Bayesian predictive uncertainty for node classification,

    M. Stadler, B. Charpentier, S. Geisler, D. Z ¨ugner, and S. G ¨unnemann, “Graph posterior network: Bayesian predictive uncertainty for node classification,” Advances in Neural Information Processing Systems , vol. 34, pp. 18 033–18 048, 2021

  14. [22]

    Com- bining graph neural networks with expert knowledge for smart contract vulnerability detection,

    Z. Liu, P. Qian, X. Wang, Y . Zhuang, L. Qiu, and X. Wang, “Com- bining graph neural networks with expert knowledge for smart contract vulnerability detection,” IEEE Transactions on Knowledge and Data Engineering, 2021

  15. [23]

    Geometric deep learning on graphs and manifolds using mixture model cnns,

    F. Monti, D. Boscaini, J. Masci, E. Rodola, J. Svoboda, and M. M. Bron- stein, “Geometric deep learning on graphs and manifolds using mixture model cnns,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 5115–5124

  16. [24]

    Neural message passing for quantum chemistry,

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, “Neural message passing for quantum chemistry,” in International conference on machine learning . PMLR, 2017, pp. 1263–1272

  17. [25]

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

    Y . Wang, Y . Liu, X. Shen, C. Li, K. Ding, R. Miao, Y . Wang, S. Pan, and X. Wang, “Unifying unsupervised graph-level anomaly de- tection and out-of-distribution detection: A benchmark,” arXiv preprint arXiv:2406.15523, 2024

  18. [26]

    Semi-supervised classification with graph convolutional networks,

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

  19. [27]

    Inductive representation learning on large graphs,

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

  20. [28]

    Towards deeper graph neural networks,

    M. Liu, H. Gao, and S. Ji, “Towards deeper graph neural networks,” in Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining , 2020, pp. 338–348

  21. [29]

    Arc: A generalist graph anomaly detector with in-context learning,

    Y . Liu, S. Li, Y . Zheng, Q. Chen, C. Zhang, and S. Pan, “Arc: A generalist graph anomaly detector with in-context learning,” in Advances in Neural Information Processing Systems , vol. 37, 2024, pp. 50 772– 50 804

  22. [30]

    Noise-resilient unsupervised graph representation learning via multi-hop feature quality estimation,

    S. Li, Y . Liu, Q. Chen, G. I. Webb, and S. Pan, “Noise-resilient unsupervised graph representation learning via multi-hop feature quality estimation,” in Proceedings of the 33rd ACM International Conference on Information and Knowledge Management , 2024, pp. 1255–1265

  23. [31]

    Predict then propagate: Graph neural networks meet personalized pagerank,

    J. Gasteiger, A. Bojchevski, and S. G ¨unnemann, “Predict then propagate: Graph neural networks meet personalized pagerank,” arXiv preprint arXiv:1810.05997, 2018

  24. [32]

    Adaptive propagation graph convolutional network,

    I. Spinelli, S. Scardapane, and A. Uncini, “Adaptive propagation graph convolutional network,” IEEE Transactions on Neural Networks and Learning Systems, vol. 32, no. 10, pp. 4755–4760, 2020

  25. [33]

    Simplifying graph convolutional networks,

    F. Wu, A. Souza, T. Zhang, C. Fifty, T. Yu, and K. Weinberger, “Simplifying graph convolutional networks,” in International conference on machine learning . PMLR, 2019, pp. 6861–6871

  26. [34]

    Sign: Scalable inception graph neural networks,

    F. Frasca, E. Rossi, D. Eynard, B. Chamberlain, M. Bronstein, and F. Monti, “Sign: Scalable inception graph neural networks,” arXiv preprint arXiv:2004.11198, 2020. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 14

  27. [35]

    Learning to drop: Robust graph neural network via topological denois- ing,

    D. Luo, W. Cheng, W. Yu, B. Zong, J. Ni, H. Chen, and X. Zhang, “Learning to drop: Robust graph neural network via topological denois- ing,” in Proceedings of the 14th ACM international conference on web search and data mining , 2021, pp. 779–787

  28. [36]

    Robust graph representation learning via neural sparsifica- tion,

    C. Zheng, B. Zong, W. Cheng, D. Song, J. Ni, W. Yu, H. Chen, and W. Wang, “Robust graph representation learning via neural sparsifica- tion,” in International Conference on Machine Learning. PMLR, 2020, pp. 11 458–11 468

  29. [37]

    Are graph convolutional networks with random weights feasible?

    C. Huang, M. Li, F. Cao, H. Fujita, Z. Li, and X. Wu, “Are graph convolutional networks with random weights feasible?” IEEE Transac- tions on Pattern Analysis and Machine Intelligence , vol. 45, no. 3, pp. 2751–2768, 2022

  30. [38]

    Auto- heg: Automated graph neural network on heterophilic graphs,

    X. Zheng, M. Zhang, C. Chen, Q. Zhang, C. Zhou, and S. Pan, “Auto- heg: Automated graph neural network on heterophilic graphs,” arXiv preprint arXiv:2302.12357, 2023

  31. [39]

    Neuromorphic camera denoising using graph neural network-driven transformers,

    Y . Alkendi, R. Azzam, A. Ayyad, S. Javed, L. Seneviratne, and Y . Zweiri, “Neuromorphic camera denoising using graph neural network-driven transformers,” IEEE Transactions on Neural Networks and Learning Systems, 2022

  32. [40]

    Interaction-aware graph neural networks for fault diagnosis of complex industrial processes,

    D. Chen, R. Liu, Q. Hu, and S. X. Ding, “Interaction-aware graph neural networks for fault diagnosis of complex industrial processes,” IEEE Transactions on neural networks and learning systems , vol. 34, no. 9, pp. 6015–6028, 2021

  33. [41]

    Guest editorial: Deep neural networks for graphs: Theory, models, algorithms, and applications,

    M. Li, A. Micheli, Y . G. Wang, S. Pan, P. Li ´o, G. S. Gnecco, and M. Sanguineti, “Guest editorial: Deep neural networks for graphs: Theory, models, algorithms, and applications,” IEEE Transactions on Neural Networks and Learning Systems , vol. 35, no. 4, pp. 4367–4372, 2024

  34. [42]

    Self-supervision im- proves diffusion models for tabular data imputation,

    Y . Liu, T. Ajanthan, H. Husain, and V . Nguyen, “Self-supervision im- proves diffusion models for tabular data imputation,” in Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, 2024, pp. 1513–1522

  35. [43]

    Goodat: towards test-time graph out-of-distribution detection,

    L. Wang, D. He, H. Zhang, Y . Liu, W. Wang, S. Pan, D. Jin, and T.-S. Chua, “Goodat: towards test-time graph out-of-distribution detection,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 14, 2024, pp. 15 537–15 545

  36. [44]

    A label-free heterophily-guided approach for unsupervised graph fraud detection,

    J. Pan, Y . Liu, X. Zheng, Y . Zheng, A. W.-C. Liew, F. Li, and S. Pan, “A label-free heterophily-guided approach for unsupervised graph fraud detection,” in Proceedings of the AAAI Conference on Artificial Intelli- gence, vol. 39, no. 12, 2025, pp. 12 443–12 451

  37. [45]

    Explainable uncertainty-aware convolutional recurrent neural network for irregular medical time series,

    Q. Tan, M. Ye, A. J. Ma, B. Yang, T. C.-F. Yip, G. L.-H. Wong, and P. C. Yuen, “Explainable uncertainty-aware convolutional recurrent neural network for irregular medical time series,” IEEE Transactions on Neural Networks and Learning Systems, vol. 32, no. 10, pp. 4665–4679, 2020

  38. [46]

    A survey of uncertainty in deep neural networks,

    J. Gawlikowski, C. R. N. Tassi, M. Ali, J. Lee, M. Humt, J. Feng, A. Kruspe, R. Triebel, P. Jung, R. Roscheret al., “A survey of uncertainty in deep neural networks,” Artificial Intelligence Review , vol. 56, no. Suppl 1, pp. 1513–1589, 2023

  39. [47]

    Correlated parameters to accurately measure uncertainty in deep neural networks,

    K. Posch and J. Pilz, “Correlated parameters to accurately measure uncertainty in deep neural networks,” IEEE Transactions on Neural Networks and Learning Systems , vol. 32, no. 3, pp. 1037–1051, 2020

  40. [48]

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning,

    Y . Gal and Z. Ghahramani, “Dropout as a bayesian approximation: Representing model uncertainty in deep learning,” in international conference on machine learning . PMLR, 2016, pp. 1050–1059

  41. [49]

    Transforming neural-net output levels to probability distributions,

    J. Denker and Y . LeCun, “Transforming neural-net output levels to probability distributions,” Advances in neural information processing systems, vol. 3, 1990

  42. [50]

    Being a bit frequentist improves bayesian neural networks,

    A. Kristiadi, M. Hein, and P. Hennig, “Being a bit frequentist improves bayesian neural networks,” in International Conference on Artificial Intelligence and Statistics . PMLR, 2022, pp. 529–545

  43. [51]

    Jøsang, Subjective logic

    A. Jøsang, Subjective logic. Springer, 2016, vol. 3

  44. [52]

    C. M. Bishop and N. M. Nasrabadi, Pattern recognition and machine learning. Springer, 2006, vol. 4, no. 4

  45. [53]

    Uncovering the structural fairness in graph contrastive learning,

    R. Wang, X. Wang, C. Shi, and L. Song, “Uncovering the structural fairness in graph contrastive learning,” Advances in Neural Information Processing Systems, vol. 35, pp. 32 465–32 473, 2022

  46. [54]

    Addressing failure prediction by learning model confidence,

    C. Corbi `ere, N. Thome, A. Bar-Hen, M. Cord, and P. P´erez, “Addressing failure prediction by learning model confidence,” Advances in Neural Information Processing Systems , vol. 32, 2019

  47. [55]

    Multimodal dynamics: Dynamical fusion for trustworthy multimodal classification,

    Z. Han, F. Yang, J. Huang, C. Zhang, and J. Yao, “Multimodal dynamics: Dynamical fusion for trustworthy multimodal classification,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 20 707–20 717

  48. [56]

    Measuring and relieving the over-smoothing problem for graph neural networks from the topological view,

    D. Chen, Y . Lin, W. Li, P. Li, J. Zhou, and X. Sun, “Measuring and relieving the over-smoothing problem for graph neural networks from the topological view,” inProceedings of the AAAI conference on artificial intelligence, vol. 34, no. 04, 2020, pp. 3438–3445

  49. [57]

    A survey on oversmooth- ing in graph neural networks,

    T. K. Rusch, M. M. Bronstein, and S. Mishra, “A survey on oversmooth- ing in graph neural networks,” arXiv preprint arXiv:2303.10993 , 2023

  50. [58]

    Dirichlet energy constrained learning for deep graph neural networks,

    K. Zhou, X. Huang, D. Zha, R. Chen, L. Li, S.-H. Choi, and X. Hu, “Dirichlet energy constrained learning for deep graph neural networks,” Advances in Neural Information Processing Systems , vol. 34, pp. 21 834–21 846, 2021

  51. [59]

    Model degradation hinders deep graph neural networks,

    W. Zhang, Z. Sheng, Z. Yin, Y . Jiang, Y . Xia, J. Gao, Z. Yang, and B. Cui, “Model degradation hinders deep graph neural networks,” in Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2022, pp. 2493–2503

  52. [60]

    Uncertainty estima- tion using a single deep deterministic neural network,

    J. Van Amersfoort, L. Smith, Y . W. Teh, and Y . Gal, “Uncertainty estima- tion using a single deep deterministic neural network,” in International conference on machine learning . PMLR, 2020, pp. 9690–9700

  53. [61]

    Confidence-aware learning for deep neural networks,

    J. Moon, J. Kim, Y . Shin, and S. Hwang, “Confidence-aware learning for deep neural networks,” in international conference on machine learning. PMLR, 2020, pp. 7034–7044

  54. [62]

    Categories of belief fusion,

    A. Jøsang, “Categories of belief fusion,” J. Adv. Inf. Fusion , vol. 13, no. 2, pp. 235–254, 2018

  55. [63]

    Uncertainty characteristics of sub- jective opinions,

    A. Josang, J.-H. Cho, and F. Chen, “Uncertainty characteristics of sub- jective opinions,” in 2018 21st International Conference on Information Fusion (FUSION). IEEE, 2018, pp. 1998–2005

  56. [64]

    Review of a mathematical theory of evidence,

    L. A. Zadeh, “Review of a mathematical theory of evidence,” AI magazine, vol. 5, no. 3, pp. 81–81, 1984

  57. [65]

    Graph attention multi-layer perceptron,

    W. Zhang, Z. Yin, Z. Sheng, Y . Li, W. Ouyang, X. Li, Y . Tao, Z. Yang, and B. Cui, “Graph attention multi-layer perceptron,” in Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2022, pp. 4560–4570

  58. [66]

    Permutation equivari- ant graph framelets for heterophilous graph learning,

    J. Li, R. Zheng, H. Feng, M. Li, and X. Zhuang, “Permutation equivari- ant graph framelets for heterophilous graph learning,”IEEE Transactions on neural networks and learning systems , 2024

  59. [67]

    Representation learning on graphs with jumping knowledge networks,

    K. Xu, C. Li, Y . Tian, T. Sonobe, K.-i. Kawarabayashi, and S. Jegelka, “Representation learning on graphs with jumping knowledge networks,” in International conference on machine learning . PMLR, 2018, pp. 5453–5462

  60. [68]

    Deepgcns: Can gcns go as deep as cnns?

    G. Li, M. Muller, A. Thabet, and B. Ghanem, “Deepgcns: Can gcns go as deep as cnns?” in Proceedings of the IEEE/CVF international conference on computer vision , 2019, pp. 9267–9276

  61. [69]

    Simple spectral graph convolution,

    H. Zhu and P. Koniusz, “Simple spectral graph convolution,” in Inter- national conference on learning representations , 2020

  62. [70]

    Agnn: Alter- nating graph-regularized neural networks to alleviate over-smoothing,

    Z. Chen, Z. Wu, Z. Lin, S. Wang, C. Plant, and W. Guo, “Agnn: Alter- nating graph-regularized neural networks to alleviate over-smoothing,” IEEE Transactions on Neural Networks and Learning Systems , 2023

  63. [71]

    Beyond message-passing: Generalization of graph neural networks via feature perturbation for semi-supervised node classification,

    Y . Choi, J. Choi, T. Ko, and C.-K. Kim, “Beyond message-passing: Generalization of graph neural networks via feature perturbation for semi-supervised node classification,” IEEE Transactions on Neural Networks and Learning Systems , 2024

  64. [72]

    Uncertainty quantification of molecular property prediction with bayesian neural networks,

    S. Ryu, Y . Kwon, and W. Y . Kim, “Uncertainty quantification of molecular property prediction with bayesian neural networks,” arXiv preprint arXiv:1903.08375, 2019

  65. [73]

    Uncertainty-aware multi-view representation learning,

    Y . Geng, Z. Han, C. Zhang, and Q. Hu, “Uncertainty-aware multi-view representation learning,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 9, 2021, pp. 7545–7553

Pith tools

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