Pith. sign in

REVIEW 3 major objections 37 references

A single scalar set by source and target homophily and feature noise tells when distribution shift makes a GNN over- or under-confident, and the exact temperature that restores calibration.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 09:07 UTC pith:P7Q72A4F

load-bearing objection Closed-form κ under CSBM is real and carefully checked; real graphs only confirm that one global T works, not that Eq. (1) predicts it. the 3 major comments →

arxiv 2607.10804 v1 pith:P7Q72A4F submitted 2026-07-12 cs.LG

When does distribution shift break graph neural networks calibration?

Abderaouf Bahi This is my paper
classification cs.LG
keywords graph neural networkscalibrationdistribution shifthomophilytemperature scalinguncertainty quantificationexpected calibration error
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper gives a closed-form account of how distribution shift moves a graph neural network’s confidence away from its accuracy. After training and calibrating on a source graph, accuracy at logit magnitude t becomes the sigmoid of κt while reported confidence remains the sigmoid of t. The slope κ is written explicitly in terms of source and target homophily and feature signal-to-noise ratio, so one can read off over-confidence, under-confidence, or perfect calibration and the single temperature that removes the gap. Under a homogeneous shift that single global temperature is population-optimal, which explains why more complex node-wise recalibrators add nothing. The resulting label-free method works on synthetic shifts but shows that estimating target accuracy without labels is the remaining barrier on real graphs.

Core claim

Under mean aggregation on a contextual stochastic block model, after source calibration, the true accuracy given logit magnitude t is σ(κt) while the model’s confidence is σ(t). The slope κ is closed-form in source and target edge homophily and the feature signal-to-noise ratio. The model is over-confident when κ < 1, under-confident when κ > 1, and the NLL-optimal temperature is exactly T⋆ = 1/κ. Under homogeneous shift this single global temperature attains minimum population NLL and zero population ECE; any non-constant per-node temperatures strictly increase NLL.

What carries the argument

The calibration slope κ = (2ht − 1)(1 + 4hs(1 − hs)ρ) / ((2hs − 1)(1 + 4ht(1 − ht)ρ)). It is the single scalar that maps reported confidence σ(t) onto true accuracy σ(κt), fixes the optimal temperature T⋆ = 1/κ, and yields the ECE bound (1/4)|κ − 1|E|δ|.

Load-bearing premise

The closed forms assume a linear Bayes-direction classifier on Gaussian features from a stochastic block model in the large-degree limit, so the aggregated signal is asymptotically Gaussian with mean proportional to (2h − 1).

What would settle it

On a homogeneous homophily shift of a CSBM, if the predicted temperature 1/κ systematically fails to match the temperature that minimizes target NLL or ECE beyond ordinary finite-degree noise, or if any non-constant per-node temperature assignment strictly lowers population NLL relative to the constant T⋆, the central claim is false.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Under homogeneous shift a single global temperature is optimal; per-node recalibrators cannot improve population NLL.
  • Source-fitted calibrators fail under shift because they lock onto κ = 1 for the source, not the target slope.
  • Covariate shift always induces over-confidence (never under-confidence), monotonically in extra feature noise.
  • One oracle temperature recovers low ECE across real graphs spanning extreme heterophily to strong homophily and multi-class settings.
  • Label-free accuracy estimation is the sole remaining bottleneck for automatic source-free GNN recalibration.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Graph-native accuracy estimators that condition on local neighborhood homophily could turn the closed-form slope into a deployable zero-label calibrator.
  • Pairing the global temperature with a conformal wrapper would combine pointwise slope correction with distribution-free set coverage.
  • The same slope viewpoint may organize other structure-dependent reliability failures, such as selective-prediction thresholds under edge rewiring.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. The paper derives a closed-form calibration slope for linear GNNs on a contextual SBM under distribution shift. After source calibration, confidence is σ(t) while accuracy is σ(κ t), with κ explicit in source/target homophily and SNR (Theorem 1, Eq. 1); over/under-confidence and the NLL-optimal temperature T⋆=1/κ follow immediately, with a matching ECE bound and extensions to self-loop GCN, multi-class, and covariate shift. Under homogeneous shift a single global temperature is population-optimal (Corollary 1). The theory is turned into STAC, a source-free label-free recalibrator that estimates target accuracy and sets one T. Synthetic checks match κ and T⋆ closely (r=0.99); on five real graphs an oracle temperature recovers ECE≤0.024, while label-free STAC is unreliable.

Significance. If the structural characterization holds beyond the generative model, this is a genuine first-principles account of how graph shift moves the confidence–accuracy gap, with falsifiable sign/magnitude predictions and a clean explanation for why global temperature scaling should suffice under homogeneous shift. Strengths include complete Appendix A proofs under stated assumptions, quantitative synthetic agreement (Fig. 8: Pearson r=0.99, MAE 0.11), an ECE bound that is never violated in simulation, and an unusually honest negative result on real-graph STAC that isolates label-free accuracy estimation as the remaining bottleneck. The operational claim that one global T recovers calibration is well supported on real data; the specific homophily/SNR formula is the more novel and load-bearing contribution.

major comments (3)
  1. The load-bearing claim is the explicit structural formula (Theorem 1 / Eq. 1, and Prop. 1 for self-loops), not merely that some global T works. Synthetic validation of that formula is strong (Fig. 8). On the five real graphs (Table 2, Fig. 10, §6.4), the paper only shows that an oracle-fitted single temperature recovers ECE ≤ 0.024. No experiment estimates (hs, ht, ρ) or the self-loop quantities A(h), B(h) and plugs them into Eq. (1)/Prop. 1 to test whether the predicted T⋆ matches the oracle. A homogeneous logit rescaling can succeed for reasons unrelated to the CSBM aggregation signal (2h−1). Please add such a plug-in test (even if approximate, with reported estimation error), or reframe the real-data claims so they support only Corollary 1’s operational consequence and not the quantitative structural characterization.
  2. §3.4 and §7.5 correctly limit the closed forms to a linear Bayes-direction head on Gaussian CSBM features in the large-degree regime (Assumptions 1–3, Lemma 1). Real experiments use SIGN/logistic features and trained 2-layer GCNs. The manuscript should state more sharply which claims are claimed to transfer (sign of over-confidence under covariate shift; sufficiency of one global T under homogeneous shift) versus which remain CSBM-only (the numerical value of κ from Eq. 1). Without that separation, the abstract’s “strong predictive power of the theory” on real graphs overstates what Table 2 actually tests.
  3. STAC (§5, Algorithm 1) is presented as a method contribution built on the theory, yet §6.4 and Table 2 show it can worsen ECE substantially (e.g. amazon-ratings 0.085 → 0.432) because ATC/disagreement overshoot accuracy. The paper is honest about this, but the contribution framing should be tightened: either (i) restrict STAC’s claimed regime to settings where the accuracy estimator is known to be reliable (synthetic / high-homophily), or (ii) treat STAC primarily as a reduction that isolates the open problem of graph-native label-free accuracy estimation (§7.4), with the oracle results as the theory’s ceiling. As written, a reader can reasonably take STAC as a deployable calibrator, which the real-graph evidence does not support.

Circularity Check

0 steps flagged

No load-bearing circularity: κ and T⋆ follow from Gaussian posteriors under CSBM aggregation plus source scale fixation; real-data oracle checks do not redefine the formula.

full rationale

The central derivation (Lemma 1 → Theorem 1 / Eq. 1 → Corollary 1) is self-contained under the stated CSBM + large-degree Gaussian assumptions. The aggregated coordinate mean (2h−1)r and variance are obtained by direct expectation/variance calculation over same-/different-class neighbours; the posterior Pr(y=+1|δ=t)=σ(κt) is the exact two-Gaussian likelihood ratio, which yields κ:=2m/s². Source calibration merely fixes the free scalar a so that κ(hs)=1; substituting produces the explicit structural formula for κ(ht). T⋆=1/κ is then the unique temperature that equates model confidence σ(t/T) to the true conditional for all t, and the population-NLL optimality of a constant temperature under homogeneous shift follows from strict convexity of NLL in the per-node inverse temperatures. None of these steps define the output in terms of itself, fit a free parameter to the target quantity being predicted, or rest on a self-citation uniqueness theorem. STAC’s accuracy estimators (ATC + disagreement) are external, off-the-shelf methods; when they fail on real graphs the paper reports the failure rather than redefining κ. The only minor self-reference is an unrelated survey citation by the same author, which is not used in any proof. Synthetic validation (r=0.99 match of predicted vs. oracle T) and real-graph oracle-temperature recovery are empirical checks, not circular redefinitions. Score 1 only for the trivial source-calibration normalisation that is explicitly declared as such.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 2 invented entities

The central claim rests on standard CSBM and large-degree Gaussian aggregation assumptions plus a source-calibrated linear Bayes-direction classifier. Free parameters appear only in the practical STAC estimator (threshold, perturbation count, optional per-node ridge), not in the closed-form κ itself. No new physical entities are postulated; κ and STAC are derived constructs.

free parameters (3)
  • source confidence threshold τ
    Carried from source validation for ATC-style accuracy estimation in Algorithm 1; not fixed by the theory of κ.
  • perturbation passes M = 6–10
    Number of input/edge perturbation passes for disagreement-based accuracy estimate; chosen as 6–10 in experiments.
  • optional per-node ridge and feature weights w,b
    Used only in the optional STAC-full stage; Corollary 1 predicts they should not help under homogeneous shift.
axioms (5)
  • domain assumption Contextual SBM with balanced classes, Gaussian features xi=μ_yi+ξi, edge homophily h=p/(p+q(K−1)), SNR ρ=r²/σ² (Assumption 1).
    Standard generative model for structure-plus-features analysis; invoked throughout Section 3–4 and Appendix A.
  • domain assumption Large-degree regime d→∞ with h,ρ fixed so aggregated coordinates are asymptotically Gaussian (Assumption 2, Lindeberg CLT).
    Enables closed-form posterior Pr(y|δ=t)=σ(κt); finite-d error O(d^{-1/2}) noted in Remark 8.
  • domain assumption Classifier is linear along the Bayes-optimal direction with scale a fixed by source calibration κ(hs)=1 (Assumption 3).
    Makes the logit a scalar multiple of the aggregated discriminant; required for Theorem 1.
  • domain assumption Regular-degree approximation for self-loop GCN normalization (Assumption 4).
    Used only for Proposition 1; error vanishes under concentrated degrees.
  • standard math Sigmoid is 1/4-Lipschitz; population ECE defined via |acc−conf| expectation.
    Standard facts used in Proposition 4 ECE bound.
invented entities (2)
  • calibration slope κ(hs,ht,ρ) independent evidence
    purpose: Single scalar that maps confidence σ(t) to accuracy σ(κt) and yields T⋆=1/κ.
    Derived from CSBM aggregation moments, not postulated as a free object; independent handle is the oracle temperature match (r=0.99) and real-graph oracle ECE recovery.
  • STAC (Structure-aware Test-time Adaptive Calibration) no independent evidence
    purpose: Source-free label-free procedure that estimates target accuracy then sets one global temperature.
    Method built on the theory; practical reliability inherits external accuracy estimators and fails on several real graphs.

pith-pipeline@v1.1.0-grok45 · 23596 in / 3827 out tokens · 47055 ms · 2026-07-14T09:07:39.639439+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of When does distribution shift break graph neural networks calibration?." pith.science (2026). https://pith.science/paper/P7Q72A4F

@misc{pith2026260710804,
  author       = {Pith},
  title        = {Pith review of: When does distribution shift break graph neural networks calibration?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P7Q72A4F}},
  note         = {Machine review of arXiv:2607.10804}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Graph neural networks (GNNs) are increasingly deployed in real-world applications where distribution shift is un-avoidable. However, how such shifts affect model calibration, defined as the agreement between predictive confidence and actual accuracy, remains poorly understood, and existing graph calibration methods typically rely on labeled validation data from the deployment distribution. In this work, I present the first closed-form theoretical characterization of GNN calibration under distribution shift. I show that calibration is governed by a single scalar quantity that explicitly depends on structural changes between the source and target graphs, as well as feature quality. This characterization precisely identifies when a model becomes over-confident, under-confident, or remains calibrated, and directly yields the optimal temperature scaling strategy. I further extend the analysis to graph convolutional networks with symmetric normalization, multi-class classification, and covariate shift, and derive a theoretical upper bound on the expected calibration error. My analysis also reveals that, under homogeneous distribution shift, a single global temperature is theoretically optimal, providing a principled explanation for why more complex node-wise recalibration methods offer no additional benefit. Building on these theoretical insights, I propose STAC, a source-free, label-free calibration method. Experiments on synthetic benchmarks demonstrate substantial calibration improvements, while evaluations on five real-world graph datasets show that reliable calibration without target labels remains challenging despite the strong predictive power of the theory.

Figures

Figures reproduced from arXiv: 2607.10804 by Abderaouf Bahi.

Figure 1
Figure 1. Figure 1: The calibration slope. A source-calibrated linear GNN reports confidence σ(t) (blue) but its true accuracy is σ(κt) (red). The shaded area is the calibration error. When homophily/covariate shift makes κ < 1 the model is over-confident; a single temperature T ⋆ = 1/κ (i.e. σ(t/T ⋆) = σ(κt)) removes the gap exactly. mechanism. I supply that mechanism for a model simple enough to analyze in closed form yet e… view at source ↗
Figure 2
Figure 2. Figure 2: Overview. Theory (top): on a CSBM, aggregation scales the class signal by (2h−1) and the noise by 1/d, giving a closed-form calibration slope κ; the model’s confidence is σ(t) but its accuracy is σ(κt), so the optimal temperature is T ⋆=1/κ. Method (middle): freeze the source-calibrated model and, on the shifted unlabeled target, estimate accuracy from perturbations to set one temperature (STAC, Section 5)… view at source ↗
Figure 3
Figure 3. Figure 3: STAC. At test time the frozen source model is applied to the shifted, unlabeled target graph; a few perturbed passes yield a label-free accuracy estimate aˆ, which fixes a single temperature T ⋆ = 1/κˆ so that mean confidence matches aˆ. No target labels and no source data are used. Remark 4. The self-loop term adds a deterministic same-class contribution (+r with certainty, since a node is always its own … view at source ↗
Figure 4
Figure 4. Figure 4: Why homophily controls calibration. Aggregation averages same- and different-class neighbours; the resulting signal along the true class is scaled by cK(h), which vanishes and flips sign at h = 1/K [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Theorem 1. Closed-form slope vs. measured 1/Toracle (left); ECE before/after T ⋆ (right). 6.3. STAC under shift (synthetic) On CSBM node classification with a trained 2-layer GCN, across feature, homophily, and heterogeneous shifts, STAC’s global stage roughly halves the ECE gap versus source-temperature scaling and approaches the oracle [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Extensions. Self-loop GCN slope κSL (left) and the K-class signal coefficient cK(h) (right): lines are theory, markers are simulation [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Proposition 4. The bound 1 4 |κ − 1| E|δ| upper-bounds the measured ECE and is tight near the source. 6.6. A different shift type: structural rewiring To test breadth beyond covariate noise I rewire a fraction of the target edges. This miscalibrates several graphs (e.g. minesweeper ECE 0.07→0.14 at 75% rewiring), and a single oracle temperature again recovers calibration across all five (ECE ≤ 0.03; [PITH… view at source ↗
Figure 8
Figure 8. Figure 8: Optimal temperature. Predicted T ⋆ = 1/κ vs. the oracle temperature across (hs, ht, ρ) configurations; points hug y=x (r=0.99, MAE 0.11). by construction, while the true target accuracy moves. Methods with a per-node correction do not escape this: a per-node function fit on source data still targets κ(hs, hs, ρ) = 1 pointwise, uniformly across nodes, so it fails in exactly the same direction as global temp… view at source ↗
Figure 9
Figure 9. Figure 9: Mechanism (synthetic feature shift). Left: ECE vs. shift intensity for the uncalibrated model, source-TS, STAC-global, and the oracle; STAC-global consistently halves the source-TS gap. Right: the true target accuracy (solid) falls as shift grows, while the source-TS assumption (dotted, frozen at the source validation accuracy) does not move at all: this divergence is the mechanism behind source-TS’s failu… view at source ↗
Figure 10
Figure 10. Figure 10: Five real graphs under covariate shift. One oracle temperature recovers calibration across the homophily spectrum; STAC is unreliable (accuracy-estimation bottleneck). currently missing. Second, the distribution-free coverage guarantees of conformalized GNNs [9] address a set-valued, not pointwise, notion of reliability; combining a globally-corrected temperature with a conformal wrapper is plausible and … view at source ↗
Figure 11
Figure 11. Figure 11: Covariate-shift sweep (Proposition 3). (a) over-confidence grows with shift strength on every real graph; (b) uncalibrated ECE (solid) rises while a single oracle temperature (dashed) stays low. Appendix A.1. Model and assumptions Assumption 1 (CSBM). There are N nodes and K balanced classes with labels yi i.i.d. uniform on {1, . . . , K}. Node features are xi = µyi +ξi , with ξi ∼ N (0, σ2 IF ) i.i.d. an… view at source ↗
Figure 12
Figure 12. Figure 12: Structural shift. Uncalibrated ECE (solid) vs. oracle temperature (dashed) as target edges are rewired; one temperature recovers calibration across graphs. The model outputs σ(δ) as the probability of class +1; for a prediction at δ = t > 0 its confidence is σ(t). Calibration requires σ(t) = σ(κt) for all t, i.e. κ = 1; the sign of κ − 1 gives over-/under-confidence. Applying temperature T maps confidence… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 4 linked inside Pith

  1. [1]

    Bahi, Graph neural networks applications across domains: All insights you need (2026).arXiv:2606.27202

    A. Bahi, Graph neural networks applications across domains: All insights you need (2026).arXiv:2606.27202. URLhttps://arxiv.org/abs/2606.27202

    Pith/arXiv arXiv 2026

  2. [2]

    X. Wang, H. Liu, C. Shi, C. Yang, Be confident! towards trustworthy graph neural networks via confidence calibration, in: Advances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  3. [3]

    H. H.-H. Hsu, Y. Shen, C. Tomani, D. Cremers, What makes graph neural networks miscalibrated?, in: Advances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  4. [4]

    C. Guo, G. Pleiss, Y. Sun, K. Q. Weinberger, On calibration of modern neural networks, in: International Conference on Machine Learning (ICML), 2017

    2017

  5. [5]

    Zhuang, C

    D. Zhuang, C. Jiang, Y. Zheng, S. Wang, J. Zhao, GETS: Ensemble temperature scaling for calibration in graph neural networks, in: International Conference on Learning Representations (ICLR), 2025

    2025

  6. [6]

    X. Li, L. Tao, H. Lu, M. Dong, J. Gao, C. Xu, WATS: Calibrating graph neural networks with wavelet-aware temperature scaling, arXiv preprint arXiv:2506.23782 (2025)

    Pith/arXiv arXiv 2025

  7. [7]

    B. Tang, Z. Wu, X. Wu, Q. Huang, J. Chen, S. Lei, H. Meng, SimCalib: Graph neural network calibration based on similarity between nodes, in: AAAI Conference on Artificial Intelligence, 2024

    2024

  8. [8]

    Trivedi, M

    P. Trivedi, M. Heimann, R. Anirudh, D. Koutra, J. J. Thiagarajan, Accurate and scalable estimation of epistemic uncertainty for graph neural networks, in: International Conference on Learning Representations (ICLR), 2024

    2024

  9. [9]

    Huang, Y

    K. Huang, Y. Jin, E. J. Candès, J. Leskovec, Uncertainty quantification over graph with conformalized graph neural networks, in: Advances in Neural Information Processing Systems (NeurIPS), 2023. 17

    2023

  10. [10]

    Y. You, T. Chen, Z. Wang, Y. Shen, Graph domain adaptation via theory-grounded spectral regularization, in: International Conference on Learning Representations (ICLR), 2023

    2023

  11. [11]

    M. Kull, M. Perello-Nieto, M. Kängsepp, T. Silva Filho, H. Song, P. Flach, Beyond temperature scaling: Ob- taining well-calibrated multi-class probabilities with dirichlet calibration, in: Advances in Neural Information Processing Systems (NeurIPS), 2019

    2019

  12. [12]

    M. P. Naeini, G. F. Cooper, M. Hauskrecht, Obtaining well calibrated probabilities using bayesian binning, in: AAAI Conference on Artificial Intelligence, 2015

    2015

  13. [13]

    Minderer, J

    M. Minderer, J. Djolonga, R. Romijnders, F. Hubis, X. Zhai, N. Houlsby, D. Tran, M. Lucic, Revisiting the calibration of modern neural networks, in: Advances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  14. [14]

    Mukhoti, V

    J. Mukhoti, V. Kulharia, A. Sanyal, S. Golodetz, P. H. S. Torr, P. K. Dokania, Calibrating deep neural networks using focal loss, in: Advances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  15. [15]

    Ovadia, E

    Y. Ovadia, E. Fertig, J. Ren, Z. Nado, D. Sculley, S. Nowozin, J. V. Dillon, B. Lakshminarayanan, J. Snoek, Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift, in: Advances in Neural Information Processing Systems (NeurIPS), 2019

    2019

  16. [16]

    M. Wang, H. Yang, Q. Cheng, GCL: Graph calibration loss for trustworthy graph neural network, in: ACM International Conference on Multimedia (ACM MM), 2022

    2022

  17. [17]

    J. Kang, Q. Zhou, H. Tong, JuryGCN: Quantifying jackknife uncertainty on graph convolutional networks, in: ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2022

    2022

  18. [18]

    Q. Wu, H. Zhang, J. Yan, D. Wipf, Handling distribution shifts on graphs: An invariance perspective, in: International Conference on Learning Representations (ICLR), 2022

    2022

  19. [19]

    Q. Zhu, N. Ponomareva, J. Han, B. Perozzi, Shift-robust GNNs: Overcoming the limitations of localized graph training data, in: Advances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  20. [20]

    S.Gui, X.Li, L.Wang, S.Ji, GOOD:Agraphout-of-distributionbenchmark, in: AdvancesinNeuralInformation Processing Systems (NeurIPS), 2022

    2022

  21. [21]

    H. Li, X. Wang, Z. Zhang, W. Zhu, OOD-GNN: Out-of-distribution generalized graph neural network, IEEE Transactions on Knowledge and Data Engineering (2022)

    2022

  22. [22]

    W. Jin, T. Zhao, J. Ding, Y. Liu, J. Tang, N. Shah, Empowering graph representation learning with test-time graph transformation, in: International Conference on Learning Representations (ICLR), 2023

    2023

  23. [23]

    D. Wang, E. Shelhamer, S. Liu, B. Olshausen, T. Darrell, Tent: Fully test-time adaptation by entropy mini- mization, in: International Conference on Learning Representations (ICLR), 2021

    2021

  24. [24]

    Y. Sun, X. Wang, Z. Liu, J. Miller, A. A. Efros, M. Hardt, Test-time training with self-supervision for general- ization under distribution shifts, in: International Conference on Machine Learning (ICML), 2020

    2020

  25. [25]

    Y. Wang, C. Li, W. Jin, R. Li, J. Zhao, J. Tang, X. Xie, Test-time training for graph neural networks, arXiv preprint arXiv:2210.08813 (2022)

    Pith/arXiv arXiv 2022

  26. [26]

    Liang, D

    J. Liang, D. Hu, J. Feng, Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation, in: International Conference on Machine Learning (ICML), 2020

    2020

  27. [27]

    S. Garg, S. Balakrishnan, Z. C. Lipton, B. Neyshabur, H. Sedghi, Leveraging unlabeled data to predict out-of- distribution performance, in: International Conference on Learning Representations (ICLR), 2022

    2022

  28. [28]

    Jiang, V

    Y. Jiang, V. Nagarajan, C. Baek, J. Z. Kolter, Assessing generalization of SGD via disagreement, in: Interna- tional Conference on Learning Representations (ICLR), 2022

    2022

  29. [29]

    D. Bo, X. Wang, C. Shi, H. Shen, Beyond low-frequency information in graph convolutional networks, in: AAAI Conference on Artificial Intelligence, 2021

    2021

  30. [30]

    Chien, J

    E. Chien, J. Peng, P. Li, O. Milenkovic, Adaptive universal generalized pagerank graph neural network, in: International Conference on Learning Representations (ICLR), 2021. 18

    2021

  31. [31]

    T. K. Rusch, B. P. Chamberlain, J. Rowbottom, S. Mishra, M. M. Bronstein, Graph-coupled oscillator networks, in: International Conference on Machine Learning (ICML), 2022

    2022

  32. [32]

    Gravina, D

    A. Gravina, D. Bacciu, C. Gallicchio, Anti-symmetric DGN: A stable architecture for deep graph networks, in: International Conference on Learning Representations (ICLR), 2023

    2023

  33. [33]

    Deshpande, S

    Y. Deshpande, S. Sen, A. Montanari, E. Mossel, Contextual stochastic block models, in: Advances in Neural Information Processing Systems (NeurIPS), 2018

    2018

  34. [34]

    T. N. Kipf, M. Welling, Semi-supervised classification with graph convolutional networks, in: International Conference on Learning Representations (ICLR), 2017

    2017

  35. [35]

    Frasca, E

    F. Frasca, E. Rossi, D. Eynard, B. Chamberlain, M. Bronstein, F. Monti, SIGN: Scalable inception graph neural networks, arXiv preprint arXiv:2004.11198 (2020)

    Pith/arXiv arXiv 2004

  36. [36]

    F. Wu, A. Souza, T. Zhang, C. Fifty, T. Yu, K. Q. Weinberger, Simplifying graph convolutional networks, in: International Conference on Machine Learning (ICML), 2019

    2019

  37. [37]

    O.Platonov, D.Kuznedelev, M.Diskin, A.Babenko, L.Prokhorenkova, AcriticallookattheevaluationofGNNs under heterophily: Are we really making progress?, in: International Conference on Learning Representations (ICLR), 2023. 19

    2023