Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

WATS: Calibrating Graph Neural Networks with Wavelet-Aware Temperature Scaling

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Heat-kernel graph wavelets, used as node-level structural signatures, can calibrate GNN confidence more accurately than one-hop neighbor statistics.

desk verdict Wavelet features for GNN calibration are a new idea worth testing, but the headline ECE claims are compromised by test-set hyperparameter selection. read the letter →

arxiv 2506.23782 v2 pith:YTUT36KB submitted 2025-06-30 cs.LG cs.AI

Xiaoyang Li , Linwei Tao , Haohui Lu , Minjing Dong , Junbin Gao , Chang Xu This is my paper
classification cs.LGcs.AI
keywords graphneuralnetworkcalibrationtemperaturescalingwaveletsheatkernelExpectedErrornode-wisepost-hocstructuralnodeembeddings
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 claims that a post-hoc calibration method which assigns each node its own temperature from heat-kernel graph wavelet features can make GNN confidence scores align much more closely with true accuracy than existing graph-aware calibrators. The authors argue that prior methods rely on one-hop neighbor statistics or latent embeddings, missing multi-hop structural context, and that wavelet coefficients provide the missing fine-grained topological signature. Across seven graph datasets and two GNN backbones, WATS reports the lowest expected calibration error among all compared methods, with gains of up to 42.3% over baselines and lower variance across runs. A sympathetic reader would care because well-calibrated GNN confidence is what lets a deployed model's probability judgments be trusted in high-stakes settings.

What carries the argument

The load-bearing object is the graph wavelet transform with a heat-kernel scaling function $g(s\lambda)=e^{-s\lambda}$, approximated by Chebyshev polynomials so that no eigendecomposition is needed. The transform is applied to the log-degree signal rather than node features, so the resulting coefficient matrix encodes each node's structural role at multiple diffusion scales; row-wise $\ell^1$ normalization yields the feature vectors. A two-layer MLP maps each node's wavelet features to a positive temperature via Softplus, and the original logits are divided by that temperature. The scale parameter $s$ controls how far diffusion reaches, and the Chebyshev order $k$ sets the receptive field, so together they tune the locality and granularity of the structural signature.

What would settle it

Run WATS on a regular graph in which all nodes have identical degree and structural role; its wavelet features then coincide and the method should reduce to a single global temperature, so it should not beat plain temperature scaling. If WATS still improves ECE in that setting, the improvement is not coming from the claimed multi-scale structural signal. Alternatively, re-run the benchmark with $k$ and $s$ chosen only from the validation set, and check whether WATS still holds the lowest ECE.

Watch

Extended reading notes

Core claim

The central claim is that the degree of miscalibration of a GNN's node-level confidence is substantially explained by the node's multi-scale structural context, and that this context can be read off from heat-kernel wavelet coefficients computed from the graph Laplacian. WATS computes these coefficients with a Chebyshev approximation of the heat-kernel scaling function applied to a log-degree signal, normalizes them, and feeds them to a two-layer MLP whose softplus output is a node-specific temperature; dividing that node's logits by this temperature recalibrates the confidence. The paper reports that this scheme achieves the lowest ECE on most dataset/backbone combinations tested, including cases where the base model is already well calibrated, and that it does so without retraining the GNN or accessing neighbor logits or predictions.

Load-bearing premise

The reported ECE advantages assume that the per-dataset wavelet hyperparameters $k$ and $s$ may be chosen using the test-set calibration error from the ablation study; if these must be selected from validation data alone, the margin over baselines could shrink or disappear.

Editorial extensions

If this is right

  • WATS can be applied to any pretrained GNN without retraining or access to neighboring logits or predictions, making it a drop-in post-hoc calibrator.
  • The wavelet features and their Chebyshev computation scale to large graphs; the reported complexity is $O(k|E| + |V|kh)$ and the transform can be precomputed once per graph.
  • On sparse or low-degree regions where one-hop statistics are weak, WATS should reduce the systematic underconfidence that GNNs exhibit.
  • Lower standard deviations across runs suggest WATS produces more stable confidence estimates than the graph-specific baselines.
  • Because the calibration signal is purely structural, the approach should transfer to other message-passing backbones beyond GCN and GAT, though the paper only tests those two.

Reading between the lines

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

  • The paper does not test whether the learned temperature mapping transfers across graphs; a reader might infer that the structural signature is generic enough to reuse, but that is an extension the paper does not claim.
  • The sensitivity analysis suggests dense graphs require precise diffusion control, so practical deployment would need a validation-based selection rule for $k$ and $s$; the paper does not provide one.
  • The same wavelet-as-uncertainty-signal idea could plausibly extend to link prediction or out-of-distribution detection, but those tasks are not evaluated here.
  • On heterophilic graphs, where one-hop labels are misleading, the multi-hop wavelet signature should help more than one-hop statistics; this follows from the paper's motivation but is not tested.
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 WATS (Wavelet-Aware Temperature Scaling), a post-hoc node-wise calibration method for GNNs. WATS computes node-specific temperatures from multi-scale structural features obtained via a Chebyshev-approximated heat-kernel graph wavelet transform, then trains a two-layer MLP on validation labels to predict temperatures and rescales logits accordingly. The authors claim that across seven benchmark datasets and two GNN backbones, WATS achieves the lowest Expected Calibration Error (ECE) among all compared methods, reduces calibration variance by 17.24% on average compared with graph-specific methods, and remains computationally efficient. The manuscript includes a preliminary study motivating multi-hop structural features, a description of the wavelet transform, an extensive experimental comparison (Table 1), an ablation study of structural features (Table 2), and a sensitivity analysis of the wavelet hyperparameters k and s (Section 4.3.2, Appendix B).

Significance. The core idea of using spatially localized, multi-scale graph wavelet features for post-hoc calibration is interesting and well motivated by the observation that one-hop statistics are insufficient to capture structural heterogeneity. If the reported results held under a fair evaluation protocol, the paper would make a useful contribution to graph neural network calibration, offering a method that is architecture-agnostic, post-hoc, and scalable. The paper includes extensive experiments on seven datasets with two backbones, and the sensitivity analysis partially documents the method's behavior. However, the main empirical claim of 'lowest ECE among all compared methods' is compromised by the use of test-set information for hyperparameter selection and is already contradicted by results in Table 1 on Cora/GAT and Pubmed/GAT.

major comments (3)
  1. [§4.3.2, Appendix A/B, Table 4] The reported WATS results in Table 1 appear to be obtained after per-dataset selection of the wavelet hyperparameters k and s using test-set ECE. Appendix A Table 4 states that k and s are 'chosen based on the Ablation study', and the ablation study in Section 4.3.2 and Tables 6–9 reports ECE over a grid of k ∈ {2,3,4,5} and s ∈ {0.1,...,2.5} on the same test splits used in Table 1. For example, Table 4 selects k=3, s=0.8 for Citeseer, and Table 7 shows the minimum ECE for Citeseer at exactly this configuration (2.15). This means the WATS numbers are effectively minima over 28 configurations per dataset, while all baseline methods are evaluated at a single configuration without equivalent test-set tuning. This makes the comparison favorable to WATS and undermines the claim that WATS 'achieves the lowest ECE among all compared methods'. Please re-evaluate using a validation-based (or nested) selection procedure for k and s, and report the effect of the selection protocol on the final ECE. The abstract should be corrected accordingly; in particular, Table 1 already shows ETS achieving 1.92 vs WATS 2.02 on Cora/GAT and TS/ETS achieving 0.88/0.87 vs WATS 0.87 on Pubmed/GAT, so the unqualified 'lowest ECE' statement is false even under the current favorable selection.
  2. [§3.3.1, Eq. (5)] The construction of the wavelet features is ambiguous. The text states that the Chebyshev polynomial approximation of the heat kernel g(sλ)=e^{-sλ} uses coefficients c_k given by the integral expression (following Hammond et al. [15]), but then defines S = Σ_{k=0}^K α_k T_k with α_k = e^{-sk} in Eq. (5). Since the Chebyshev coefficients of e^{-sλ} are not equal to e^{-sk}, Eq. (5) is not the stated approximation. Please clarify whether α_k should be the true coefficients c_k, or provide a justification for the choice α_k = e^{-sk}. As written, the exact wavelet feature transform is not well-defined and cannot be reproduced from the paper (code is not yet released).
  3. [Abstract and §4] The abstract claims that WATS reduces calibration variance by 17.24% on average compared with graph-specific methods, but this statistic does not appear in the main text or any table. The only variance-related information is the standard deviations in Table 1, which do not by themselves support this specific quantitative claim. Please provide the variance-reduction analysis (e.g., averaged standard deviations across runs or datasets) or remove the claim.
minor comments (4)
  1. [§3.2, Eq. (3)] The one-hop bias expression and the following example are inconsistent: if the estimator includes the self-loop, then for di=2 and neighbor labels [0,1], the average is (y_i+1)/3, which depends on y_i, contrary to the statement that 'the average is 1/3 regardless of the true label yi'. Please revise the equation or the accompanying text for clarity.
  2. [§4.4] The complexity analysis omits the cost of precomputing the wavelet features for a given graph; this is presumably a one-time cost and should be clarified, since it affects the practical efficiency comparison with methods that do not require such precomputation.
  3. [§1 and throughout] There are several typographical issues, including inconsistent spacing in 'W ATS' and grammatical slips such as 'Differs from graph wavelet algorithms used in neural network'. A careful proofreading pass is recommended.
  4. [Table 1] The caption states that 'Best performance on ECE are highlighted for each configuration', but in the plain-text version it is not evident which entries are highlighted; please ensure the final PDF clearly marks the best result in each row.

Circularity Check

1 steps flagged · score 6.0 of 10

WATS's headline ECE superiority is partially circular: k and s are selected using test-set ECE from the ablation study, so the reported 'lowest ECE' is a test-set-informed selection rather than an independent prediction; the core wavelet-to-temperature derivation is otherwise self-contained.

  1. fitted input called prediction [Section 4.1 Calibration Settings; Section 4.3.2; Appendix A (Table 4); Table 1]
    "For each method, calibration parameters are learned on the validation set and evaluated on the test set. ... The hyperparameter of graph wavelet k and s are chosen based on the Ablation study. ... To assess the robustness of W ATS, we perform an exhaustive grid search over the Chebyshev order k ∈ {2, 3, 4, 5} and the heat-kernel scale s ∈ {0.1, 0.4, 0.8, 1.2, 1.6, 2.0, 2.5} on seven node-classification benchmarks, we visualize the changes of ECE for varying k and s ... (full results about hyperparameters are in Appendix)."

    The reported WATS ECE values in Table 1 use per-dataset k and s from Table 4, and Appendix A states these are 'chosen based on the Ablation study' (Section 4.3.2, Tables 6-9). Since ECE is evaluated on the test set (Section 4.1), the WATS numbers are test-set-informed selections rather than the output of a fixed method; baselines receive no comparable test-set hyperparameter tuning. This makes the headline 'lowest ECE' claim partially a selected optimum, not an independent prediction. The core step--MLP temperature trained on validation labels and evaluated on test labels--remains non-circular.

full rationale

One partial circularity affects the headline empirical claim: the per-dataset wavelet hyperparameters (k,s) are selected using the test-set ECE ablation (Section 4.3.2, Tables 6-9; Appendix A), so the WATS numbers in Table 1 are test-informed selections and the comparison against baselines without equivalent test tuning is biased. This reduces the 'lowest ECE' claim to the selection procedure. However, the method's derivation itself is self-contained: wavelet features are computed from the graph Laplacian and log-degree; the MLP temperature is trained on validation cross-entropy; test ECE is then measured honestly. The only self-citation [36] is a general statement about GNN underconfidence and is not load-bearing. The inconsistency between alpha_k=e^{-sk} and the Chebyshev coefficient c_k is a correctness/ambiguity issue, not circularity.

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

The central claim depends on two hand-tuned wavelet hyperparameters selected with test-set information, and on the empirical assumption that log-degree-driven wavelet features capture calibration-relevant structure. No new entities are proposed.

free parameters (4)
  • Chebyshev order k = 2 to 4 per dataset
    Chosen by hand based on test-set ablation (Appendix B); controls receptive field size.
  • Heat-kernel scale s = 0.4 to 1.6 per dataset
    Chosen by hand based on test-set ablation (Appendix B); controls diffusion extent.
  • MLP hidden dimension = 16 to 128 per dataset
    Chosen by hand (Table 4); part of the temperature predictor architecture.
  • MLP dropout = 0.2 to 0.95 per dataset
    Chosen by hand (Table 4); regularization for the temperature predictor.
assumptions (4)
  • standard math Chebyshev polynomial expansion approximates the heat-kernel wavelet operator
    Invoked in Sec. 3.3.1 to avoid eigendecomposition; relies on Hammond et al.
  • domain assumption GNNs are systematically underconfident on low-degree nodes
    Motivation in Sec. 3.2 and Figure 1; used to justify structure-aware calibration.
  • ad hoc to paper Wavelet features computed from log-degree input capture structural uncertainty better than one-hop statistics
    Core modeling assumption; no theoretical proof, only empirical ablation in Table 2.
  • ad hoc to paper alpha_k = e^{-sk} is a valid Chebyshev coefficient approximation
    Eq. (5) uses this without derivation from the stated integral formula in Sec 3.3.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WATS: Calibrating Graph Neural Networks with Wavelet-Aware Temperature Scaling." pith.science (2026). https://pith.science/paper/YTUT36KB

@misc{pith2026250623782,
  author       = {Pith},
  title        = {Pith review of: WATS: Calibrating Graph Neural Networks with Wavelet-Aware Temperature Scaling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YTUT36KB}},
  note         = {Machine review of arXiv:2506.23782}
}
read the original abstract

Graph Neural Networks (GNNs) have demonstrated strong predictive performance on relational data; however, their confidence estimates often misalign with actual predictive correctness, posing significant limitations for deployment in safety-critical settings. While existing graph-aware calibration methods seek to mitigate this limitation, they primarily depend on coarse one-hop statistics, such as neighbor-predicted confidence, or latent node embeddings, thereby neglecting the fine-grained structural heterogeneity inherent in graph topology. In this work, we propose Wavelet-Aware Temperature Scaling (WATS), a post-hoc calibration framework that assigns node-specific temperatures based on tunable heat-kernel graph wavelet features. Specifically, WATS harnesses the scalability and topology sensitivity of graph wavelets to refine confidence estimates, all without necessitating model retraining or access to neighboring logits or predictions. Extensive evaluations across seven benchmark datasets with varying graph structures and two GNN backbones demonstrate that WATS achieves the lowest Expected Calibration Error (ECE) among all compared methods, outperforming both classical and graph-specific baselines by up to 42.3\% in ECE and reducing calibration variance by 17.24\% on average compared with graph-specific methods. Moreover, WATS remains computationally efficient, scaling well across graphs of diverse sizes and densities. Code will be released based on publication.

Figures

Figures reproduced from arXiv: 2506.23782 by the authors.

Figure 1
Figure 1. The x-axis shows node degree ranges and the y-axis represents the ECE. The blue bars [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. In both plots, “Uncali” refers to the uncalibrated model and “Cali” refers to the calibrated [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Sensitivity analysis of wavelet hyper-parameters. Each plot shows the ECE scores on [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Calibration performance of Cora dataset. [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Calibration performance of Computers dataset. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Calibration performance of Photo dataset. [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: Calibration performance of Cora-full dataset. [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Calibration performance of Reddit dataset. [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: Calibration performance of Pubmed dataset. [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]

Discussion (0). Sign in 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. When does distribution shift break graph neural networks calibration?

    cs.LG 2026-07 conditional novelty 7.0 of 10

    GNN calibration under distribution shift is governed by a single closed-form slope κ(hs, ht, ρ) that sets the optimal global temperature T⋆=1/κ and explains when node-wise recalibration cannot help.

Reference graph

Works this paper leans on

54 extracted references · 36 canonical work pages · cited by 1 Pith paper

  1. [15]

    Wavelets on graphs via spectral graph theory

    David K Hammond, Pierre Vandergheynst, and Rémi Gribonval. Wavelets on graphs via spectral graph theory. Applied and Computational Harmonic Analysis, 30(2):129–150, 2011

    work page 2011

  2. [1]

    Geometric multimodal deep learning with multiscaled graph wavelet convolutional network

    Maysam Behmanesh, Peyman Adibi, Sayyed Mohammad Saeed Ehsani, and Jocelyn Chanussot. Geometric multimodal deep learning with multiscaled graph wavelet convolutional network. IEEE Transactions on Neural Networks and Learning Systems, 2022

    work page 2022

  3. [2]

    Deep gaussian embedding of graphs: Unsu- pervised inductive learning via ranking

    Aleksandar Bojchevski and Stephan Günnemann. Deep gaussian embedding of graphs: Unsu- pervised inductive learning via ranking. arXiv preprint arXiv:1707.03815, 2017

    arXiv 2017

  4. [3]

    Bruce, D

    A. Bruce, D. Donoho, and H.-Y . Gao. Wavelet analysis [for signal processing].IEEE Spectrum, 33(10):26–35, 1996. doi: 10.1109/6.540087

    work page doi:10.1109/6.540087 1996

  5. [4]

    Graph wavelets for spatial traffic analysis

    Mark Crovella and Eric Kolaczyk. Graph wavelets for spatial traffic analysis. In IEEE INFOCOM 2003. Twenty-second Annual Joint Conference of the IEEE Computer and Commu- nications Societies (IEEE Cat. No. 03CH37428), volume 3, pages 1848–1857. IEEE, 2003

    work page 2003

  6. [5]

    The laplacian spectrum of a graph

    K Ch Das. The laplacian spectrum of a graph. Computers & Mathematics with Applications, 48 (5-6):715–724, 2004

    work page 2004

  7. [6]

    Sea-gwnn: Simple and effective adaptive graph wavelet neural network

    Swakshar Deb, Sejuti Rahman, and Shafin Rahman. Sea-gwnn: Simple and effective adaptive graph wavelet neural network. Proceedings of the AAAI Conference on Artificial Intelligence, 38(10):11740–11748, Mar. 2024. doi: 10.1609/aaai.v38i10.29058. URL https://ojs.aaai. org/index.php/AAAI/article/view/29058

    doi:10.1609/aaai.v38i10.29058 2024

  8. [7]

    Learning structural node embeddings via diffusion wavelets

    Claire Donnat, Marinka Zitnik, David Hallac, and Jure Leskovec. Learning structural node embeddings via diffusion wavelets. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1320–1329, 2018

    2018

Show all 54 references
  1. [8]

    Graph neural networks for social recommendation, 2019

    Wenqi Fan, Yao Ma, Qing Li, Yuan He, Eric Zhao, Jiliang Tang, and Dawei Yin. Graph neural networks for social recommendation, 2019. URL https://arxiv.org/abs/1902.07243

    2019 arXiv

  2. [9]

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

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. Ininternational conference on machine learning, pages 1050–1059. PMLR, 2016

    2016

  3. [10]

    Medical- knowledge-based graph neural network for medication combination prediction

    Chao Gao, Shu Yin, Haiqiang Wang, Zhen Wang, Zhanwei Du, and Xuelong Li. Medical- knowledge-based graph neural network for medication combination prediction. IEEE Transactions on Neural Networks and Learning Systems , 35(10):13246–13257, 2024. doi: 10.1109/TNNLS.2023.3266490

    2024

  4. [11]

    Citeseer: An automatic citation indexing system

    C Lee Giles, Kurt D Bollacker, and Steve Lawrence. Citeseer: An automatic citation indexing system. In Proceedings of the third ACM conference on Digital libraries, pages 89–98, 1998

    1998

  5. [12]

    Gaunt, Alvaro Sanchez-Gonzalez, Yulia Rubanova, Petar Velickovic, James Kirkpatrick, and Peter W

    Jonathan Godwin, Michael Schaarschmidt, Alexander L. Gaunt, Alvaro Sanchez-Gonzalez, Yulia Rubanova, Petar Velickovic, James Kirkpatrick, and Peter W. Battaglia. Very deep graph neural networks via noise regularisation. CoRR, abs/2106.07971, 2021. URL https: //arxiv.org/abs/2106.07971

    2021 arXiv

  6. [13]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. CoRR, abs/1706.04599, 2017. URL http://arxiv.org/abs/1706.04599

    2017 arXiv

  7. [14]

    Inductive representation learning on large graphs

    Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017

    2017

  8. [16]

    What makes graph neural networks miscalibrated? Advances in Neural Information Processing Systems, 35: 13775–13786, 2022

    Hans Hao-Hsun Hsu, Yuesong Shen, Christian Tomani, and Daniel Cremers. What makes graph neural networks miscalibrated? Advances in Neural Information Processing Systems, 35: 13775–13786, 2022

    2022

  9. [17]

    Semi-supervised classification with graph convolutional networks

    Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 11

    2016 arXiv

  10. [18]

    Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers

    Meelis Kull, Telmo Silva Filho, and Peter Flach. Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers. In Artificial intelligence and statistics, pages 623–631. PMLR, 2017

    2017

  11. [19]

    On calibration of graph neural networks for node classification

    Tong Liu, Yushan Liu, Marcel Hildebrandt, Mitchell Joblin, Hang Li, and V olker Tresp. On calibration of graph neural networks for node classification. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2022

    2022

  12. [20]

    A weighted patient network-based framework for predicting chronic diseases using graph neural networks

    Haohui Lu and Shahadat Uddin. A weighted patient network-based framework for predicting chronic diseases using graph neural networks. Scientific reports, 11(1):22607, 2021

    2021

  13. [21]

    Cross-links matter for link prediction: Rethinking the debiased gnn from a data perspective

    Zihan Luo, Hong Huang, Jianxun Lian, Xiran Song, Xing Xie, and Hai Jin. Cross-links matter for link prediction: Rethinking the debiased gnn from a data perspective. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neu- ral Informatio...

    2023

  14. [22]

    Bayesian neural networks and density networks

    David J.C MacKay. Bayesian neural networks and density networks. Nuclear Instru- ments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detec- tors and Associated Equipment , 354(1):73–80, 1995. ISSN 0168-9002. doi: https://doi. org/10.1016/0168-9002(94)...

    1995

  15. [23]

    Automating the construction of internet portals with machine learning

    Andrew Kachites McCallum, Kamal Nigam, Jason Rennie, and Kristie Seymore. Automating the construction of internet portals with machine learning. Information Retrieval, 3:127–163, 2000

    2000

  16. [24]

    Bird, Diego R

    Gledson Melotti, Cristiano Premebida, Jordan J. Bird, Diego R. Faria, and Nuno Gonçalves. Reducing overconfidence predictions in autonomous driving perception. IEEE Access, 10: 54805–54821, 2022. doi: 10.1109/ACCESS.2022.3175195

    2022

  17. [25]

    Resnikoff and Raymond O

    Howard L. Resnikoff and Raymond O. Wells. Wavelet analysis and the geometry of euclidean domains. Journal of Geometry and Physics , 8(1):273–282, 1992. ISSN 0393-0440. doi: https://doi.org/10.1016/0393-0440(92)90052-3. URL https://www.sciencedirect.com/ science/article/pii/039...

    1992

  18. [26]

    Conformalized quantile regression

    Yaniv Romano, Evan Patterson, and Emmanuel Candes. Conformalized quantile regression. Advances in neural information processing systems, 32, 2019

    2019

  19. [27]

    Collective classification in network data

    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

    2008

  20. [28]

    A graph neural network (gnn)-based approach for real-time estimation of traffic speed in sustainable smart cities

    Amit Sharma, Ashutosh Sharma, Polina Nikashina, Vadim Gavrilenko, Alexey Tselykh, Alexander Bozhenyuk, Mehedi Masud, and Hossam Meshref. A graph neural network (gnn)-based approach for real-time estimation of traffic speed in sustainable smart cities. Sustainability, 15(15), 2...

    2023 doi

  21. [29]

    Pitfalls of graph neural network evaluation

    Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. Pitfalls of graph neural network evaluation. Relational Representation Learning Workshop, NeurIPS 2018, 2018

    2018

  22. [30]

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

    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 Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 2270–2279, 2023

    2023

  23. [31]

    The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains

    David I Shuman, Sunil K Narang, Pascal Frossard, Antonio Ortega, and Pierre Vandergheynst. The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains. IEEE signal processing magazine, 30(3):83–98, 2013. 12

    2013

  24. [32]

    Bayesian optimization with robust bayesian neural networks

    Jost Tobias Springenberg, Aaron Klein, Stefan Falkner, and Frank Hutter. Bayesian optimization with robust bayesian neural networks. Advances in neural information processing systems, 29, 2016

    2016

  25. [33]

    Generalized equivariance and preferential labeling for gnn node classification

    Zeyu Sun, Wenjie Zhang, Lili Mou, Qihao Zhu, Yingfei Xiong, and Lu Zhang. Generalized equivariance and preferential labeling for gnn node classification. Proceedings of the AAAI Conference on Artificial Intelligence, 36(8):8395–8403, Jun. 2022. doi: 10.1609/aaai.v36i8. 20815. ...

    2022 doi

  26. [34]

    The lifting scheme: A construction of second generation wavelets

    Wim Sweldens. The lifting scheme: A construction of second generation wavelets. SIAM journal on mathematical analysis, 29(2):511–546, 1998

    1998

  27. [35]

    Simcalib: Graph neural network calibration based on similarity between nodes

    Boshi Tang, Zhiyong Wu, Xixin Wu, Qiaochu Huang, Jun Chen, Shun Lei, and Helen Meng. Simcalib: Graph neural network calibration based on similarity between nodes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 15267–15275, 2024

    2024

  28. [36]

    Feature clipping for uncertainty calibration

    Linwei Tao, Minjing Dong, and Chang Xu. Feature clipping for uncertainty calibration. Proceedings of the AAAI Conference on Artificial Intelligence , 39(19):20841–20849, Apr

  29. [37]

    Conformal prediction under covariate shift

    Ryan J Tibshirani, Rina Foygel Barber, Emmanuel Candes, and Aaditya Ramdas. Conformal prediction under covariate shift. Advances in neural information processing systems, 32, 2019

    2019

  30. [38]

    Graph wavelets for multiscale community mining

    Nicolas Tremblay and Pierre Borgnat. Graph wavelets for multiscale community mining. IEEE Transactions on Signal Processing, 62(20):5227–5239, 2014

    2014

  31. [39]

    Graph attention networks

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017

    2017 arXiv

  32. [40]

    Gcl: Graph calibration loss for trustworthy graph neural network

    Min Wang, Hao Yang, and Qing Cheng. Gcl: Graph calibration loss for trustworthy graph neural network. In Proceedings of the 30th ACM International Conference on Multimedia , pages 988–996, 2022

    2022

  33. [41]

    Be confident! towards trustworthy graph neural networks via confidence calibration

    Xiao Wang, Hongrui Liu, Chuan Shi, and Cheng Yang. Be confident! towards trustworthy graph neural networks via confidence calibration. Advances in Neural Information Processing Systems, 34:23768–23779, 2021

    2021

  34. [42]

    Mitigating overconfidence in large language models: A behavioral lens on confidence estimation and calibration

    Bingbing Wen, Chenjun Xu, Robert Wolfe, Lucy Lu Wang, Bill Howe, et al. Mitigating overconfidence in large language models: A behavioral lens on confidence estimation and calibration. In NeurIPS 2024 Workshop on Behavioral Machine Learning, 2024

    2024

  35. [43]

    Graph wavelet neural network

    Bingbing Xu, Huawei Shen, Qi Cao, Yunqi Qiu, and Xueqi Cheng. Graph wavelet neural network. arXiv preprint arXiv:1904.07785, 2019

    1904 arXiv

  36. [44]

    Cali- brating graph neural networks from a data-centric perspective

    Cheng Yang, Chengdong Yang, Chuan Shi, Yawen Li, Zhiqiang Zhang, and Jun Zhou. Cali- brating graph neural networks from a data-centric perspective. In Proceedings of the ACM Web Conference 2024, pages 745–755, 2024

    2024

  37. [45]

    Balanced confidence calibration for graph neural networks

    Hao Yang, Min Wang, Qi Wang, Mingrui Lao, and Yun Zhou. Balanced confidence calibration for graph neural networks. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3747–3757, 2024

    2024

  38. [46]

    Learning and making decisions when costs and probabili- ties are both unknown

    Bianca Zadrozny and Charles Elkan. Learning and making decisions when costs and probabili- ties are both unknown. In Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining, pages 204–213, 2001

    2001

  39. [47]

    Transforming classifier scores into accurate multiclass probability estimates

    Bianca Zadrozny and Charles Elkan. Transforming classifier scores into accurate multiclass probability estimates. In Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 694–699, 2002

    2002

  40. [48]

    Mix-n-match: Ensemble and compositional methods for uncertainty calibration in deep learning

    Jize Zhang, Bhavya Kailkhura, and T Yong-Jin Han. Mix-n-match: Ensemble and compositional methods for uncertainty calibration in deep learning. In International conference on machine learning, pages 11117–11128. PMLR, 2020. 13

    2020

  41. [49]

    Link prediction based on graph neural networks

    Muhan Zhang and Yixin Chen. Link prediction based on graph neural networks. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, ed- itors, Advances in Neural Information Processing Systems , volume 31. Curran Associates, Inc., 2018. URL https:...

    2018

  42. [50]

    Motif-based graph self-supervised learning for molecular property prediction

    Zaixi Zhang, Qi Liu, Hao Wang, Chengqiang Lu, and Chee-Kong Lee. Motif-based graph self-supervised learning for molecular property prediction. In M. Ranzato, A. Beygelz- imer, Y . Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neu- ral Information Processing...

    2021

  43. [51]

    Graphsmote: Imbalanced node classification on graphs with graph neural networks

    Tianxiang Zhao, Xiang Zhang, and Suhang Wang. Graphsmote: Imbalanced node classification on graphs with graph neural networks. InProceedings of the 14th ACM International Conference on Web Search and Data Mining , WSDM ’21, page 833–841, New York, NY , USA, 2021. Association f...

    2021

  44. [52]

    How framelets enhance graph neural networks

    Xuebin Zheng, Bingxin Zhou, Junbin Gao, Yu Guang Wang, Pietro Lió, Ming Li, and Guido Montúfar. How framelets enhance graph neural networks. arXiv preprint arXiv:2102.06986, 2021

    2021 arXiv

  45. [53]

    Gets: Ensemble temperature scaling for calibration in graph neural networks

    Dingyi Zhuang, Chonghe Jiang, Yunhan Zheng, Shenhao Wang, and Jinhua Zhao. Gets: Ensemble temperature scaling for calibration in graph neural networks. arXiv preprint arXiv:2410.09570, 2024. 14 A Experiment setting We randomly conduct the train test split 10 times for each dat...

    2024 arXiv

  46. [2025]

    URL https://ojs.aaai.org/index.php/AAAI/ article/view/34297

    doi: 10.1609/aaai.v39i19.34297. URL https://ojs.aaai.org/index.php/AAAI/ article/view/34297

    doi

Pith tools

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