Pith. sign in

REVIEW 3 major objections 6 minor 106 references

Effects of Dropout on Performance in Long-range Graph Learning Tasks

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read Random edge-dropping methods make deep graph networks less sensitive to distant nodes and hurt their performance on long-range tasks; the paper introduces DropSens to restore that sensitivity.

desk verdict Solid theory on DropEdge-variants shrinking L-hop sensitivity, but the paper overclaims the causal story—DropNode is the tell. read the letter →

arxiv 2502.07364 v2 pith:W4U3BKXE submitted 2025-02-11 cs.LG

classification cs.LG
keywords graphneuralnetworksover-squashingDropEdgelong-rangelearningsensitivityanalysisdropoutrewiringmessagepassing
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

Deep graph networks that pass messages locally suffer from over-squashing: information from far-away nodes gets crushed before it can influence a node's representation. This paper argues that the popular random-dropping techniques used to make deep networks trainable, such as DropEdge, DropAgg, DropGNN, and dropout-style variants, make over-squashing worse because on average they weaken every edge that carries information across long distances. In a linear graph convolutional network the paper proves this precisely: with asymmetric normalization, any of these methods lowers the expected sensitivity of a node to features of nodes $L$ hops away, and the sensitivity falls further as the dropping probability rises. The same conclusion is extended to nonlinear networks through existing sensitivity bounds and Monte Carlo influence estimates, and confirmed on synthetic and real-world long-range benchmarks, where the best dropout configurations usually underperform the no-drop baseline. To fix the problem the paper introduces DropSens, which chooses per-edge dropping probabilities so that the fraction of information preserved along each incoming edge stays fixed; on GCNs, DropSens beats graph-rewiring methods designed to reduce over-squashing on node-classification tasks.

What carries the argument

The engine of the argument is the expected propagation matrix $\dot P := \mathbb{E}[\hat A]$ under the random edge masks, together with the identity that expected sensitivity in a linear GCN equals $\bigl(\mathbb{E}[\hat A]^L\bigr)_{ij}\lVert W\rVert_1$. Because masks are independent across layers, expectation passes through the product, so the $L$-step transition probability of this random walk controls how much node $i$'s output depends on node $j$'s input. For DropEdge, Lemma 3.1 gives $\dot P_{ii} = (1-q^{d_i+1})/((1-q)(d_i+1))$ and $\dot P_{ij} = (1-\dot P_{ii})/d_i$, with analogous entries for DropAgg and DropGNN that have even larger self-loop weights. The proof of Theorem 3.1 exploits that a path from $j$ to $i$ with $d_G(j,i)=L$ contains no self-loop, so every factor on the path is strictly smaller than in the no-drop case. DropSens inverts this by fixing the fraction $c$ of information preserved over a cross-edge through $1-c = (q_i - q_i^{d_i+1})/(d_i(1-q_i))$, producing per-target-degree dropping probabilities that keep distant edges stronger.

What would settle it

On the SyntheticZINC benchmark at a high commute-time percentile, train 11-layer GCNs with DropEdge at $q=0.2$ and $q=0.5$, and compute the mean expected sensitivity $\bigl(\mathbb{E}[\hat A]^L\bigr)_{ij}$ for the selected distant pairs together with test MAE; the paper's claim implies sensitivity falls and MAE rises as $q$ increases, so observing higher sensitivity with better MAE, or lower sensitivity with worse MAE, would show that expected sensitivity is not the controlling factor.

Watch

Extended reading notes

Core claim

The central discovery is stated as Theorem 3.1: in an $L$-layer linear GCN with $\hat A = \hat A_{\mathrm{asym}}$, using DropEdge, DropAgg, or DropGNN decreases the sensitivity of a node $i$ to a node $j$ in $S^{(L)}(i)$, thereby reducing the effective receptive field, and the sensitivity decreases monotonically as the dropping probability increases. The reason is that the expected propagation matrix $\mathbb{E}[\hat A]$ under these masks behaves like a non-uniform random walk: its self-loop entries become larger and its cross-edge entries become smaller than in a no-drop model. Every path connecting nodes exactly $L$ hops apart contains no self-loops, so each factor in the expansion of $(\mathbb{E}[\hat A]^L)_{ij}$ shrinks, and with it the expected Jacobian norm that measures over-squashing. The paper derives exact expected matrix entries for DropEdge (Lemma 3.1), gives analogous entries for DropAgg and DropGNN, extends existing nonlinear MPNN sensitivity upper bounds to random masks, and empirically verifies exponential decay of expected sensitivity with distance on the Cora graph. On long-range benchmarks, the best-performing dropout configurations typically fail to beat NoDrop, while DropSens—which keeps a fixed information fraction $c$ on each incoming edge—raises distant sensitivity and outperforms rewiring baselines on GCN node classification while remaining competitive on graph classification.

Load-bearing premise

The load-bearing premise is that the expected sensitivity of random-mask message passing, computed while ignoring the optimization trajectory, is what determines whether a dropout-trained GNN will generalize well on a long-range task; if training dynamics override this signal, the predicted connection to test accuracy weakens.

Editorial extensions

If this is right

  • On homophilic, short-range datasets, dropout-style methods improve test accuracy in the majority of configurations, but on heterophilic, long-range datasets they usually underperform the no-drop baseline; evaluations that only use short-range tasks therefore overstate the general usefulness of these methods.
  • Increasing the dropping probability monotonically reduces expected sensitivity between nodes $L$ hops apart, so the same knob that relieves over-smoothing also worsens over-squashing.
  • DropSens, which holds the per-edge information fraction fixed, improves distant sensitivity and outperforms graph-rewiring baselines on GCN node-classification tasks while using the same number of dropped edges and fewer propagated messages.
  • Graph-level long-range tasks are less affected by dropping methods, since the readout mixes information from distant nodes; this suggests over-squashing has limited impact on graph-level performance.
  • The theoretical guarantees cover degree-normalized GCN-like message passing; GAT, ResGCN, and graph transformers fall outside the analysis, and DropSens itself does not transfer to GIN's message-passing scheme.

Reading between the lines

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

  • We infer that the expected-sensitivity diagnostic could be computed before training on any fixed graph, providing a cheap way to predict whether a given dropout probability is safe for a long-range task without running the training loop.
  • We infer that the per-edge information-preservation principle behind DropSens could be applied to feature dropout or attention dropout by using the actual induced edge weights under the mask; the paper does not test this extension.
  • We infer that benchmark suites for methods intended to train deep GNNs should routinely include heterophilic and synthetic commute-time tasks, because homophilic-only evaluations cannot separate local fitting from long-range capacity.
  • We infer that architectures with global propagation, such as graph transformers, may need no DropSens-like correction, whereas attention-based local message passing could benefit from a sensitivity-aware masking analog.
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 / 6 minor

Summary. This paper studies the effect of random-dropping methods (DropEdge, DropNode, DropAgg, DropGNN, Dropout, DropMessage) on long-range graph learning. In a linear GCN with asymmetric normalization, the authors compute the expected propagation matrix under each edge-dropping variant (Lemma 3.1 and Appendix B.1) and prove Theorem 3.1 that DropEdge, DropAgg, and DropGNN reduce expected sensitivity between nodes at distance exactly L, with sensitivity decreasing in the dropping probability. They also extend nonlinear sensitivity bounds to the random-dropping setting (Appendix C.1). Based on this analysis, they propose DropSens, which sets per-node dropping probabilities to preserve a target fraction c of information across edges. Experiments on SyntheticZINC, homophilic/heterophilic node classification, and graph classification show that dropping methods generally improve homophilic and harm heterophilic performance, and DropSens improves over DropEdge on heterophilic tasks and is competitive with graph-rewiring baselines on several datasets.

Significance. The paper addresses a real gap: dropout-style methods for deep GNNs have mostly been evaluated on short-range tasks, and their effect on over-squashing is underexplored. The theoretical analysis is clean and, within its stated scope (linear GCN, asymmetric normalization, exact L-hop pairs), the Lemma and Theorem appear correct. The paper also provides reproducible experiments and a practical algorithm, DropSens, with implementations in the appendix. However, the central causal claim, that the proved sensitivity decay explains observed long-range performance degradation, is undermined by the paper's own DropNode results and by the explicit decision to ignore optimization effects, which the experiments show are necessary to explain DropNode's behavior. The manuscript is valuable but requires substantial revision to align its claims with what is actually proven.

major comments (3)
  1. [Section 3, Eq. (B.15), Table 1a, Appendix F.2] The abstract and conclusion state that 'DropEdge-variants reduce sensitivity between distant nodes, limiting their suitability for long-range tasks,' and that these variants 'exacerbate the over-squashing problem.' However, Eq. (B.15) shows that DropNode's expected propagation matrix is exactly the NoDrop matrix, so Theorem 3.1 does not predict any sensitivity loss for DropNode. Table 1a nevertheless shows DropNode consistently degrading heterophilic accuracy (e.g., GCN on Chameleon -0.674, on Squirrel -0.656; GAT on Chameleon -6.090), and Appendix F.2 explicitly attributes this degradation to 'the effects on the learning trajectory,' an effect the theory brackets in Section 3. Thus, for one member of the named family, the observed long-range degradation is not caused by the theoretical mechanism the paper proves; the causal story requires the optimization trajectory to do decisive work. This is an internal inconsistency in the central claim and should be resolved by either excluding DropNode from the theoretical statements or by providing a separate, testable mechanism for its behavior.
  2. [Section 3, Appendix E.2] The theory computes the expected sensitivity of the stochastic network under random masks while explicitly ignoring the optimization trajectory. At test time, the dropping methods are turned off (q=0), as stated in Appendix E.2: 'isolating the effects on optimization and generalization, which our theory does not address.' Nevertheless, the paper uses the sensitivity decay to explain test-time performance differences; for example, Section 5.2 concludes that the dropping methods 'harm generalization in long-range tasks by forcing models to overfit to short-range signals.' This is a causal leap: Theorem 3.1 characterizes the training-time expected sensitivity at initialization, not the generalization behavior of the learned parameters. The DropNode case is a concrete counterexample in which the empirical degradation has no sensitivity-based explanation. To support the central claim, the paper should either directly measure sensitivity of the trained models (e.g., on test-time graphs) and correlate it with performance, or substantially weaken the causal language connecting Theorem 3.1 to the experimental results.
  3. [Section 3, Theorem 3.1, Abstract] The theorem is proved only for node pairs at distance exactly L in an L-layer GCN, and the text explicitly states that no general conclusion follows for nodes separated by fewer than L hops or for arbitrary distances. The abstract, however, states without qualification that 'theoretical results showing that DropEdge-variants reduce sensitivity between distant nodes' and the conclusion repeats this broader claim. The empirical Figure 1a does support a broader decay for DropEdge on Cora, but the theoretical contribution is narrower. The claims should be aligned with the proven scope, or the theory should be extended to all distances.
minor comments (6)
  1. [Section 2.2, Abstract] The paper uses 'DropEdge-variants' to include DropNode, DropAgg, and DropGNN, while Dropout and DropMessage are called 'Dropout-variants'; however, the theoretical results in Section 3 only cover edge-dropping methods, and the abstract's phrase 'DropEdge- and Dropout-variants exacerbate the over-squashing problem' goes beyond the theory for Dropout and DropMessage.
  2. [Section 4, Eq. (4.1)] The statement in the abstract that DropSens increases sensitivity to distant nodes 'despite dropping the same number of edges' is ambiguous; DropSens drops edges according to a target-node degree distribution, so the expected number of dropped edges differs from DropEdge for a given q. Please clarify whether 'same number' refers to the expected total.
  3. [Section 5.2] The best-performing dropping probability is selected from up to nine values per dataset-model-dropout combination after 20 runs, and then a t-test is performed on 50 new samples; no correction for multiple comparisons is applied, which can inflate the reported significance.
  4. [Figure 1b] 'relatively insignificant' should be 'relatively small' to avoid the statistical meaning of 'insignificant.'
  5. [Table 2b, Abstract] DropSens does not consistently outperform the rewiring baselines on graph classification (e.g., Mutag 70.20 vs. FoSR 80.00 and DIGL 79.70; Proteins 70.61 vs. FoSR 73.42), so the abstract's claim of consistently outperforming graph rewiring techniques is too strong; Section 5.3's qualified 'performs competitively' should be used in the abstract as well.
  6. [Appendix E.3, Listing 1] The expression sympy.real_roots(...)[-2] is fragile; please document the ordering of the roots and add a comment explaining why the second-to-last root is the desired solution.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem 3.1 follows from a first-principles expectation calculation, DropSens is a construction rather than a prediction, and no self-citation chain or fitted-input-as-prediction step appears.

full rationale

The paper's central theoretical claim is derived self-containedly rather than assumed. Equation 3.2 expresses expected sensitivity as (E[Â])^L_{ij} ||W||_1 under i.i.d. layer-wise edge masks; Lemma 3.1 and Appendix B.1 compute E[Â] directly from the Bernoulli mask distributions for each DropEdge variant; Appendix B.2 proves Theorem 3.1 by comparing the product of expected transition probabilities along loop-free L-hop paths with the corresponding NoDrop entries. No fitted parameter is later renamed as a prediction, and the theorem is stated for fixed weights and graph topology. The DropSens construction in Equation 4.1 is also not circular: it sets a per-edge dropping probability so that a user-chosen fraction c of cross-edge information is preserved, so the resulting increase in expected sensitivity is a design guarantee, not a fitted explanation of the experimental results. The paper does not use the DropSens benchmark outcomes to tune c or to retrofit the theory. There is no load-bearing self-citation: the cited sensitivity and rewiring results [8, 33, 44, 79] are by other authors, and the paper invokes no uniqueness theorem from its own prior work. The manuscript does contain an acknowledged internal limitation, flagged explicitly in Appendix F.2: DropNode has no expected sensitivity loss (Equation B.15), yet its test performance degrades on heterophilic datasets, and the paper attributes this to learning-trajectory effects that Section 3 says it is ignoring. This weakens the abstract's blanket statement that DropEdge-variants reduce sensitivity between distant nodes, but it is a correctness and scope concern, not a circular reduction of an output to an input.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical or mathematical entities are postulated; DropSens is an algorithm, not an entity. The main free parameters are the dropout probabilities and DropSens's c and q_max, which are chosen per dataset in the experiments. The axioms reflect the theory's restriction to linear networks, layer-wise independent masks, and the decision to ignore optimization effects.

free parameters (3)
  • DropSens information-preservation fraction c = chosen from {0.5, 0.8, 0.9, 0.95} per dataset
    Controls the target fraction of information preserved over cross-edges in Eq. 4.1; selected per task using validation performance, so it is a fitted/choice parameter for the empirical DropSens results.
  • DropSens maximum dropping probability q_max = chosen from {0.2, 0.3, 0.5, 0.8} per dataset
    Caps the per-edge dropping probabilities and filters configurations that would reduce to plain DropEdge; selected via validation as part of the 11-configuration search.
  • Dropout probability q per method = grid {0.1, ..., 0.9}, best per dataset
    For each dropping method, the best q is selected on validation from nine values before reporting test accuracy, so reported comparative results depend on this per-dataset selection.
assumptions (4)
  • domain assumption Asymmetric normalization A_hat = D_tilde^{-1} A_tilde for the theoretical sensitivity analysis
    Invoked in Section 3 before Lemma 3.1 to make the expected propagation matrix tractable; the paper handles symmetric normalization only through Monte Carlo approximation in Appendix D.1.
  • standard math Edge masks are independent across layers, giving E[product of A_hat^(l)] = (E[A_hat])^L
    Used in Eq. 3.2 and Theorem 3.1; this follows from the standard layer-wise independent sampling of DropEdge variants, though within a layer the mask entries are correlated through degree normalization.
  • ad hoc to paper Expected sensitivity under random masks, ignoring optimization effects, predicts test-time generalization on long-range tasks
    Stated explicitly in Section 3: 'we will ignore the effect of DropEdge-variants on the optimization trajectory'; the paper's main empirical conclusions map the sensitivity result onto test performance, making this the load-bearing bridge.
  • domain assumption The Jacobian-norm sensitivity definition of [79] is the operative measure of over-squashing
    Adopted from prior work and used as the central object of the theoretical analysis, so the validity of the over-squashing statements is tied to that definition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Effects of Dropout on Performance in Long-range Graph Learning Tasks." pith.science (2026). https://pith.science/paper/W4U3BKXE

@misc{pith2026250207364,
  author       = {Pith},
  title        = {Pith review of: Effects of Dropout on Performance in Long-range Graph Learning Tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W4U3BKXE}},
  note         = {Machine review of arXiv:2502.07364}
}
read the original abstract

Message Passing Neural Networks (MPNNs) are a class of Graph Neural Networks (GNNs) that propagate information across the graph via local neighborhoods. The scheme gives rise to two key challenges: over-smoothing and over-squashing. While several Dropout-style algorithms, such as DropEdge and DropMessage, have successfully addressed over-smoothing, their impact on over-squashing remains largely unexplored. This represents a critical gap in the literature, as failure to mitigate over-squashing would make these methods unsuitable for long-range tasks -- the intended use case of deep MPNNs. In this work, we study the aforementioned algorithms, and closely related edge-dropping algorithms -- DropNode, DropAgg and DropGNN -- in the context of over-squashing. We present theoretical results showing that DropEdge-variants reduce sensitivity between distant nodes, limiting their suitability for long-range tasks. To address this, we introduce DropSens, a sensitivity-aware variant of DropEdge that explicitly controls the proportion of information lost due to edge-dropping, thereby increasing sensitivity to distant nodes despite dropping the same number of edges. Our experiments on long-range synthetic and real-world datasets confirm the predicted limitations of existing edge-dropping and feature-dropping methods. Moreover, DropSens consistently outperforms graph rewiring techniques designed to mitigate over-squashing, suggesting that simple, targeted modifications can substantially improve a model's ability to capture long-range interactions. Our conclusions highlight the need to re-evaluate and re-design existing methods for training deep GNNs, with a renewed focus on modelling long-range interactions.

Figures

Figures reproduced from arXiv: 2502.07364 by the authors.

Figure 1
Figure 1. Empirical sensitivity analysis using the Cora dataset. 1. increases the sensitivity of a node’s representations to its own input features, and 2. decreases the sensitivity to its neighbors’ features. L-layer Linear GCNs. Unfortunately, we cannot draw similar conclusions in L-layer networks, for nodes at arbitrary distances. To see this, view P˙ as the transition matrix of a non-uniform random walk. This walk has hig… view at source ↗
Figure 2
Figure 2. Train and test MAE of 11-layer GCNs on the SyntheticZINC dataset, averaged over 10 initializations. 4 Sensitivity-Aware DropEdge Lemma 3.1 tells us that DropEdge decreases the weight of cross-edges, (j → i), in the expected propagation matrix, i.e. the strength of message passing over these edges decreases. The fraction of information preserved over a cross-edge is dependent only on the dropping probability and the … view at source ↗
Figure 3
Figure 3. Relative change in test-time performance of a GCN using DropSens, compared to the baseline DropEdge, on real-world datasets from Section 5.2. In this work, we use Cora [57], CiteSeer [31] and PubMed [62] as representatives of homophilic datasets [52, 98], and Squirrel, Chameleon and TwitchDE [69] to represent heterophilic datasets [52]. The networks’ statistics are presented in [PITH_FULL_IMAGE:figures/full_fig_p00… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Entries of P¨6 , averaged after binning node-pairs by their shortest distance [PITH_FULL_IMAGE:figures/full_fig_p025_4.png]
Figure 5
Figure 5. Figure 5: Entries of P6 ℓ=0 P˙ ℓ , averaged after binning node-pairs by their shortest distance. D.1 Symmetrically Normalized Propagation Matrix The results in Section 3 correspond to the use of Aˆ = Aˆasym for aggregating messages – in each message passing step, only the in-deg…
Figure 6
Figure 6. Figure 6: Edge-wise dropping probabilities under DropSens for varying values of c, along with corresponding approximations as in Equation E.1. mapper [ d_i ] = q_i in_degrees = degrees [ edge_index [1]] # Edge index -> in - degree of target node qs = mapper [ in_degrees ] # Edge…
Figure 7
Figure 7. Figure 7: Dropping probability versus test accuracy of DropEdge-GCN. The theory the explains the contrasting trends as follows: random edge-dropping pushes models to fit to local information during training, which is suitable for short-range tasks, but harms test-time performanc…
Figure 8
Figure 8. Figure 8: Dropping probability versus test accuracy of DropNode-GCN. F.2 Remark on DropNode In Equation B.15, we noted that DropNode does not suffer from loss in sensitivity. However, those results were in expectation. Moreover, our analysis did not account for the effects on th…
Figure 9
Figure 9. Figure 9: DropEdge probability versus training accuracy of GCNs. The training performance improves with q, suggesting that the models are not underfitting. Instead, the reason for poor test-time performance ( [PITH_FULL_IMAGE:figures/full_fig_p031_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

106 extracted references · 75 canonical work pages

  1. [1]

    Expanders via local edge flips

    Zeyuan Allen-Zhu, Aditya Bhaskara, Silvio Lattanzi, Vahab Mirrokni, and Lorenzo Orecchia. Expanders via local edge flips. InProceedings of the Twenty-Seventh Annual ACM-SIAM Sym- posium on Discrete Algorithms, SODA ’16, pp. 259–269, USA, 2016. Society for Industrial and Applied Mathematics

  2. [2]

    Alon and V

    N. Alon and V . D. Milman.λ1, isoperimetric inequalities for graphs, and superconcentrators. Journal of Combinatorial Theory, Series B, 38(1):73–88, February 1985

  3. [3]

    Eigenvalues and expanders.Combinatorica, 6(2):83–96, June 1986

    Noga Alon. Eigenvalues and expanders.Combinatorica, 6(2):83–96, June 1986

  4. [4]

    On the bottleneck of graph neural networks and its practical impli- cations

    Uri Alon and Eran Yahav. On the bottleneck of graph neural networks and its practical impli- cations. InInternational Conference on Learning Representations, 2021

  5. [5]

    Diffwire: Inductive graph rewiring via the lov ´asz bound

    Adri ´an Arnaiz-Rodr´ıguez, Ahmed Begga, Francisco Escolano, and Nuria M Oliver. Diffwire: Inductive graph rewiring via the lov ´asz bound. In Bastian Rieck and Razvan Pascanu (eds.), Proceedings of the First Learning on Graphs Conference, volume 198 ofProceedings of Ma- chine Learning Research, pp. 15:1–15:27. PMLR, 12 2022

  6. [6]

    Banerjee, Kedar Karhadkar, Yu Guang Wang, Uri Alon, and Guido Mont ´ufar

    Pradeep Kr. Banerjee, Kedar Karhadkar, Yu Guang Wang, Uri Alon, and Guido Mont ´ufar. Oversquashing in gnns through the lens of information contraction and graph expansion. In 2022 58th Annual Allerton Conference on Communication, Control, and Computing (Allerton), pp. 1–8. IEEE Press, 2022

  7. [7]

    Kostylev, Mikael Monet, Jorge P ´erez, Juan Reutter, and Juan Pablo Silva

    Pablo Barcel ´o, Egor V . Kostylev, Mikael Monet, Jorge P ´erez, Juan Reutter, and Juan Pablo Silva. The logical expressiveness of graph neural networks. InInternational Conference on Learning Representations, 2020

  8. [8]

    Understanding oversquash- ing in GNNs through the lens of effective resistance

    Mitchell Black, Zhengchao Wan, Amir Nayyeri, and Yusu Wang. Understanding oversquash- ing in GNNs through the lens of effective resistance. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett (eds.),Proceedings of the 40th International Conference on Machine Learning, volume 202 ofProceedings of Machin...

Show all 106 references
  1. [9]

    Borgwardt, Cheng Soon Ong, Stefan Sch ¨onauer, S

    Karsten M. Borgwardt, Cheng Soon Ong, Stefan Sch ¨onauer, S. V . N. Vishwanathan, Alex J. Smola, and Hans-Peter Kriegel. Protein function prediction via graph kernels.Bioinformatics, 21(suppl 1):i47–i56, 06 2005

  2. [10]

    Chandra, Prabhakar Raghavan, Walter L

    Ashok K. Chandra, Prabhakar Raghavan, Walter L. Ruzzo, Roman Smolensky, and Prasoon Ti- wari. The electrical resistance of a graph captures its commute and cover times.computational complexity, 6:312–340, 1989

  3. [11]

    Stochastic training of graph convolutional networks with variance reduction

    Jianfei Chen, Jun Zhu, and Le Song. Stochastic training of graph convolutional networks with variance reduction. InInternational Conference on Machine Learning, pp. 941–949, 2018

  4. [12]

    FastGCN: Fast learning with graph convolutional net- works via importance sampling

    Jie Chen, Tengfei Ma, and Cao Xiao. FastGCN: Fast learning with graph convolutional net- works via importance sampling. InInternational Conference on Learning Representations, 2018

  5. [13]

    Simple and deep graph convolutional networks

    Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. Simple and deep graph convolutional networks. In Hal Daum ´e III and Aarti Singh (eds.),Proceedings of the 37th International Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Researc...

  6. [14]

    Lawrence Erlbaum Associates, Hillsdale, NJ, 2 edition, 1988

    Jacob Cohen.Statistical Power Analysis for the Behavioral Sciencesstical power analysis for the behavioral sciences. Lawrence Erlbaum Associates, Hillsdale, NJ, 2 edition, 1988

  7. [15]

    The flip markov chain for connected regular graphs.Discrete Applied Mathematics, 254:56–79, 2019

    Colin Cooper, Martin Dyer, Catherine Greenhill, and Andrew Handley. The flip markov chain for connected regular graphs.Discrete Applied Mathematics, 254:56–79, 2019

  8. [16]

    Expander graph propagation

    Andreea Deac, Marc Lackenby, and Petar Veliˇckovi´c. Expander graph propagation. InNeurIPS 2022 Workshop: New Frontiers in Graph Learning, 2022. 10

  9. [17]

    Lopez de Compadre, Gargi Debnath, Alan J

    Asim Kumar Debnath, Rosa L. Lopez de Compadre, Gargi Debnath, Alan J. Shusterman, and Corwin Hansch. Structure-activity relationship of mutagenic aromatic and heteroaromatic nitro compounds. correlation with molecular orbital energies and hydrophobicity.Journal of Medicinal Ch...

  10. [18]

    On over-squashing in message passing neural networks: The impact of width, depth, and topology

    Francesco Di Giovanni, Lorenzo Giusti, Federico Barbero, Giulia Luise, Pietro Lio, and Michael M Bronstein. On over-squashing in message passing neural networks: The impact of width, depth, and topology. InInternational Conference on Machine Learning, pp. 7865–

  11. [19]

    Dobson and Andrew J

    Paul D. Dobson and Andrew J. Doig. Distinguishing enzyme structures from non-enzymes without alignments.Journal of Molecular Biology, 330(4):771–783, 2003

  12. [20]

    Long range graph benchmark

    Vijay Prakash Dwivedi, Ladislav Ramp ´aˇsek, Mikhail Galkin, Ali Parviz, Guy Wolf, Anh Tuan Luu, and Dominique Beaini. Long range graph benchmark. InThirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022

  13. [21]

    Dropmessage: Unifying random dropping for graph neural networks.Proceedings of the AAAI Conference on Artificial Intelligence, 37(4):4267–4275, Jun

    Taoran Fang, Zhiqing Xiao, Chunping Wang, Jiarong Xu, Xuan Yang, and Yang Yang. Dropmessage: Unifying random dropping for graph neural networks.Proceedings of the AAAI Conference on Artificial Intelligence, 37(4):4267–4275, Jun. 2023

  14. [22]

    A local switch markov chain on given degree graphs with application in connectivity of peer-to-peer networks

    Tomas Feder, Adam Guetz, Milena Mihail, and Amin Saberi. A local switch markov chain on given degree graphs with application in connectivity of peer-to-peer networks. In2006 47th Annual IEEE Symposium on Foundations of Computer Science (FOCS’06), pp. 69–76, 2006

  15. [23]

    Graph random neural networks for semi-supervised learning on graphs

    Wenzheng Feng, Jie Zhang, Yuxiao Dong, Yu Han, Huanbo Luan, Qian Xu, Qiang Yang, Evgeny Kharlamov, and Jie Tang. Graph random neural networks for semi-supervised learning on graphs. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (eds.),Advances in Neural Inf...

  16. [24]

    Matthias Fey and Jan E. Lenssen. Fast graph representation learning with PyTorch Geometric. InICLR Workshop on Representation Learning on Graphs and Manifolds, 2019

  17. [25]

    Rewiring with positional encodings for graph neural networks.Transactions on Machine Learning Research, 2023

    Rickard Br ¨uel Gabrielsson, Mikhail Yurochkin, and Justin Solomon. Rewiring with positional encodings for graph neural networks.Transactions on Machine Learning Research, 2023

  18. [26]

    Bayesian convolutional neural networks with bernoulli approximate variational inference, 2016

    Yarin Gal and Zoubin Ghahramani. Bayesian convolutional neural networks with bernoulli approximate variational inference, 2016

  19. [27]

    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. In Maria Florina Balcan and Kilian Q. Weinberger (eds.),Pro- ceedings of The 33rd International Conference on Machine Learning, volume 48 ofProceed- ings of M...

  20. [28]

    A theoretically grounded application of dropout in re- current neural networks

    Yarin Gal and Zoubin Ghahramani. A theoretically grounded application of dropout in re- current neural networks. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett (eds.),Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc., 2016

  21. [29]

    Large-scale learnable graph convolutional networks

    Hongyang Gao, Zhengyang Wang, and Shuiwang Ji. Large-scale learnable graph convolutional networks. InProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 1416–1424. ACM, 2018

  22. [30]

    Expanders via local edge flips in quasilinear time

    George Giakkoupis. Expanders via local edge flips in quasilinear time. InProceedings of the 54th Annual ACM SIGACT Symposium on Theory of Computing, STOC 2022, pp. 64–76, New York, NY , USA, 2022. Association for Computing Machinery

  23. [31]

    Lee Giles, Kurt D

    C. Lee Giles, Kurt D. Bollacker, and Steve Lawrence. Citeseer: an automatic citation indexing system. InProceedings of the Third ACM Conference on Digital Libraries, DL ’98, pp. 89–98, New York, NY , USA, 1998. Association for Computing Machinery. 11

  24. [32]

    Schoenholz, Patrick F

    Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural message passing for quantum chemistry. In Doina Precup and Yee Whye Teh (eds.), Proceedings of the 34th International Conference on Machine Learning, volume 70 ofPro- ceedings of M...

  25. [33]

    Konstantin Rusch, Michael Bronstein, Andreea Deac, Marc Lack- enby, Siddhartha Mishra, and Petar Veliˇckovi´c

    Francesco Di Giovanni, T. Konstantin Rusch, Michael Bronstein, Andreea Deac, Marc Lack- enby, Siddhartha Mishra, and Petar Veliˇckovi´c. How does over-squashing affect the power of GNNs?Transactions on Machine Learning Research, 2024

  26. [34]

    Giraldo, Konstantinos Skianis, Thierry Bouwmans, and Fragkiskos D

    Jhony H. Giraldo, Konstantinos Skianis, Thierry Bouwmans, and Fragkiskos D. Malliaros. On the trade-off between over-smoothing and over-squashing in deep graph neural networks. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management, CIKM...

  27. [35]

    DRew: Dynamically rewired message passing with delay

    Benjamin Gutteridge, Xiaowen Dong, Michael M Bronstein, and Francesco Di Giovanni. DRew: Dynamically rewired message passing with delay. InInternational Conference on Machine Learning, pp. 12252–12267. PMLR, 2023

  28. [36]

    Inductive representation learning on large graphs

    Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.),Advances in Neural Information Processing Systems, volume 30. Curran Associa...

  29. [37]

    Bayesian graph neural networks with adaptive con- nection sampling, 2020

    Arman Hasanzadeh, Ehsan Hajiramezanali, Shahin Boluki, Mingyuan Zhou, Nick Duffield, Krishna Narayanan, and Xiaoning Qian. Bayesian graph neural networks with adaptive con- nection sampling, 2020

  30. [38]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016

  31. [39]

    Larry V . Hedges. Distribution theory for glass’s estimator of effect size and related estima- tors.Journal of Educational Statistics, 6(2):107–128, 2025/03/15/ 1981. Full publication date: Summer, 1981

  32. [40]

    Open graph benchmark: Datasets for machine learning on graphs

    Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. arXiv preprint arXiv:2005.00687, 2020

  33. [41]

    Normalization techniques in training dnns: Methodology, analysis and application, 2020

    Lei Huang, Jie Qin, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Normalization techniques in training dnns: Methodology, analysis and application, 2020

  34. [42]

    ZINC: a free tool to discover chemistry for biology.J Chem Inf Model, 52(7):1757–1768, June 2012

    John J Irwin, Teague Sterling, Michael M Mysinger, Erin S Bolstad, and Ryan G Coleman. ZINC: a free tool to discover chemistry for biology.J Chem Inf Model, 52(7):1757–1768, June 2012

  35. [43]

    Dropagg: Robust graph neural networks via drop aggregation.Neural Networks, 163:65–74, 2023

    Bo Jiang, Yong Chen, Beibei Wang, Haiyun Xu, and Bin Luo. Dropagg: Robust graph neural networks via drop aggregation.Neural Networks, 163:65–74, 2023

  36. [44]

    Banerjee, and Guido Montufar

    Kedar Karhadkar, Pradeep Kr. Banerjee, and Guido Montufar. FoSR: First-order spectral rewiring for addressing oversquashing in GNNs. InThe Eleventh International Conference on Learning Representations, 2023

  37. [45]

    Deep learning without poor local minima

    Kenji Kawaguchi. Deep learning without poor local minima. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett (eds.),Advances in Neural Information Processing Sys- tems, volume 29. Curran Associates, Inc., 2016

  38. [46]

    Adam: A method for stochastic optimization

    Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InInterna- tional Conference on Learning Representations (ICLR), San Diega, CA, USA, 2015

  39. [47]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. InInternational Conference on Learning Representations (ICLR), 2017. 12

  40. [48]

    Howard, Wayne Hubbard, and Lawrence Jackel

    Yann LeCun, Bernhard Boser, John Denker, Donnie Henderson, R. Howard, Wayne Hubbard, and Lawrence Jackel. Handwritten digit recognition with a back-propagation network. In D. Touretzky (ed.),Advances in Neural Information Processing Systems, volume 2. Morgan- Kaufmann, 1989

  41. [49]

    Deepgcns: Can gcns go as deep as cnns? In2019 IEEE/CVF International Conference on Computer Vision (ICCV), pp

    Guohao Li, Matthias M ¨uller, Ali Thabet, and Bernard Ghanem. Deepgcns: Can gcns go as deep as cnns? In2019 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 9266–9275, 2019

  42. [50]

    Deeper insights into graph convolutional networks for semi-supervised learning.Proceedings of the AAAI Conference on Artificial Intelligence, 32(1), 04 2018

    Qimai Li, Zhichao Han, and Xiao-ming Wu. Deeper insights into graph convolutional networks for semi-supervised learning.Proceedings of the AAAI Conference on Artificial Intelligence, 32(1), 04 2018

  43. [51]

    Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard S. Zemel. Gated graph sequence neural networks. In Yoshua Bengio and Yann LeCun (eds.),4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings, 2016

  44. [52]

    New benchmarks for learning on non-homophilous graphs.arXiv preprint arXiv:2104.01404, 2021

    Derek Lim, Xiuyu Li, Felix Hohne, and Ser-Nam Lim. New benchmarks for learning on non-homophilous graphs.arXiv preprint arXiv:2104.01404, 2021

  45. [53]

    Towards deeper graph neural networks

    Meng Liu, Hongyang Gao, and Shuiwang Ji. Towards deeper graph neural networks. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. ACM, 2020

  46. [54]

    Curvdrop: A ricci curvature based approach to prevent graph neural networks from over- smoothing and over-squashing

    Yang Liu, Chuan Zhou, Shirui Pan, Jia Wu, Zhao Li, Hongyang Chen, and Peng Zhang. Curvdrop: A ricci curvature based approach to prevent graph neural networks from over- smoothing and over-squashing. InProceedings of the ACM Web Conference 2023, WWW ’23, pp. 221–230, New York, ...

  47. [55]

    Lov ´asz

    L. Lov ´asz. Random walks on graphs: A survey.Combinatorics, Paul Erdos is Eighty, 2(1): 1–46, 1993

  48. [56]

    Peer-to-peer networks based on random trans- formations of connected regular undirected graphs

    Peter Mahlmann and Christian Schindelhauer. Peer-to-peer networks based on random trans- formations of connected regular undirected graphs. InProceedings of the Seventeenth Annual ACM Symposium on Parallelism in Algorithms and Architectures, SPAA ’05, pp. 155–164, New York, NY...

  49. [57]

    Automating the construction of internet portals with machine learning.Information Retrieval, 3(2):127– 163, 07 2000

    Andrew Kachites McCallum, Kamal Nigam, Jason Rennie, and Kristie Seymore. Automating the construction of internet portals with machine learning.Information Retrieval, 3(2):127– 163, 07 2000

  50. [58]

    Wiki-cs: A wikipedia-based benchmark for graph neural networks.arXiv preprint arXiv:2007.02901, 2020

    P ´eter Mernyei and C˘at˘alina Cangea. Wiki-cs: A wikipedia-based benchmark for graph neural networks.arXiv preprint arXiv:2007.02901, 2020

  51. [59]

    Smith, Mateusz Paprocki, Ond ˇrej ˇCert´ık, Sergey B

    Aaron Meurer, Christopher P. Smith, Mateusz Paprocki, Ond ˇrej ˇCert´ık, Sergey B. Kirpichev, Matthew Rocklin, AMiT Kumar, Sergiu Ivanov, Jason K. Moore, Sartaj Singh, Thilina Rath- nayake, Sean Vig, Brian E. Granger, Richard P. Muller, Francesco Bonazzi, Harsh Gupta, Shivam V...

  52. [60]

    Geometric matrix completion with recurrent multi-graph neural networks

    Federico Monti, Michael Bronstein, and Xavier Bresson. Geometric matrix completion with recurrent multi-graph neural networks. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.),Advances in Neural Information Processing Syste...

  53. [61]

    Kriege, Franka Bause, Kristian Kersting, Petra Mutzel, and Mar- ion Neumann

    Christopher Morris, Nils M. Kriege, Franka Bause, Kristian Kersting, Petra Mutzel, and Mar- ion Neumann. Tudataset: A collection of benchmark datasets for learning with graphs. In ICML 2020 Workshop on Graph Representation Learning and Beyond (GRL+ 2020), 2020

  54. [62]

    Query-driven active sur- veying for collective classification

    Galileo Mark Namata, Ben London, Lise Getoor, and Bert Huang. Query-driven active sur- veying for collective classification. InInternational Workshop on Mining and Learning with Graphs, Edinburgh, Scotland, 2012. MLG. 13

  55. [63]

    Revis- iting over-smoothing and over-squashing using ollivier-ricci curvature

    Khang Nguyen, Hieu Nong, Vinh Nguyen, Nhat Ho, Stanley Osher, and Tan Nguyen. Revis- iting over-smoothing and over-squashing using ollivier-ricci curvature. InProceedings of the 40th International Conference on Machine Learning, ICML’23. JMLR.org, 2023

  56. [64]

    Graph neural networks exponentially lose expressive power for node classification

    Kenta Oono and Taiji Suzuki. Graph neural networks exponentially lose expressive power for node classification. InInternational Conference on Learning Representations, 2020

  57. [65]

    Dropgnn: Random dropouts increase the expressiveness of graph neural networks

    P ´al Andr´as Papp, Karolis Martinkus, Lukas Faber, and Roger Wattenhofer. Dropgnn: Random dropouts increase the expressiveness of graph neural networks. In M. Ranzato, A. Beygelz- imer, Y . Dauphin, P.S. Liang, and J. Wortman Vaughan (eds.),Advances in Neural Information Proc...

  58. [66]

    Geom-gcn: Geometric graph convolutional networks

    Hongbin Pei, Bingzhe Wei, Kevin Chen-Chuan Chang, Yu Lei, and Bo Yang. Geom-gcn: Geometric graph convolutional networks. InInternational Conference on Learning Represen- tations, 2020

  59. [67]

    Probabilistically rewired message-passing neural networks

    Chendi Qian, Andrei Manolache, Kareem Ahmed, Zhe Zeng, Guy Van den Broeck, Mathias Niepert, and Christopher Morris. Probabilistically rewired message-passing neural networks. InThe Twelfth International Conference on Learning Representations, 2024

  60. [68]

    Dropedge: Towards deep graph convolutional networks on node classification

    Yu Rong, Wenbing Huang, Tingyang Xu, and Junzhou Huang. Dropedge: Towards deep graph convolutional networks on node classification. InInternational Conference on Learning Representations, 2020

  61. [69]

    Multi-Scale Attributed Node Embedding

    Benedek Rozemberczki, Carl Allen, and Rik Sarkar. Multi-Scale Attributed Node Embedding. Journal of Complex Networks, 9(2), 2021

  62. [70]

    Konstantin Rusch, Michael M

    T. Konstantin Rusch, Michael M. Bronstein, and Siddhartha Mishra. A survey on oversmooth- ing in graph neural networks, 2023

  63. [71]

    The graph neural network model.IEEE Transactions on Neural Networks, 20(1):61–80, 2009

    Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfar- dini. The graph neural network model.IEEE Transactions on Neural Networks, 20(1):61–80, 2009

  64. [72]

    Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling

    Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. Modeling relational data with graph convolutional networks, 2017

  65. [73]

    BRENDA, the enzyme database: updates and major new developments.Nucleic Acids Res, 32(Database issue):D431–3, January 2004

    Ida Schomburg, Antje Chang, Christian Ebeling, Marion Gremse, Christian Heldt, Gregor Huhn, and Dietmar Schomburg. BRENDA, the enzyme database: updates and major new developments.Nucleic Acids Res, 32(Database issue):D431–3, January 2004

  66. [74]

    S. S. Shapiro and M. B. Wilk. An analysis of variance test for normality (complete samples). Biometrika, 52(3/4):591–611, 2025/03/12/ 1965. Full publication date: Dec., 1965

  67. [75]

    Pitfalls of graph neural network evaluation.Relational Representation Learning Workshop, NeurIPS 2018, 2018

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

  68. [76]

    Approximate counting, uniform generation and rapidly mixing markov chains.Information and Computation, 82(1):93–133, 1989

    Alistair Sinclair and Mark Jerrum. Approximate counting, uniform generation and rapidly mixing markov chains.Information and Computation, 82(1):93–133, 1989

  69. [77]

    Dropout: A simple way to prevent neural networks from overfitting.Journal of Machine Learning Research, 15(56):1929–1958, 2014

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdi- nov. Dropout: A simple way to prevent neural networks from overfitting.Journal of Machine Learning Research, 15(56):1929–1958, 2014

  70. [78]

    Going deeper with convolutions

    Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1–9, 2015

  71. [79]

    Bronstein

    Jake Topping, Francesco Di Giovanni, Benjamin Paul Chamberlain, Xiaowen Dong, and Michael M. Bronstein. Understanding over-squashing and bottlenecks on graphs via curva- ture. InInternational Conference on Learning Representations, 2022. 14

  72. [80]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.),Advances in Neural In...

  73. [81]

    Graph attention networks

    Petar Veli ˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li `o, and Yoshua Bengio. Graph attention networks. InInternational Conference on Learning Rep- resentations, 2018

  74. [82]

    Comparison of descriptor spaces for chemical compound retrieval and classification

    Nikil Wale and George Karypis. Comparison of descriptor spaces for chemical compound retrieval and classification. InSixth International Conference on Data Mining (ICDM’06), pp. 678–689, 2006

  75. [83]

    Representing long-range context for graph neural networks with global attention

    Zhanghao Wu, Paras Jain, Matthew Wright, Azalia Mirhoseini, Joseph E Gonzalez, and Ion Stoica. Representing long-range context for graph neural networks with global attention. In Advances in Neural Information Processing Systems (NeurIPS), 2021

  76. [84]

    Representation learning on graphs with jumping knowledge networks

    Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichi Kawarabayashi, and Stefanie Jegelka. Representation learning on graphs with jumping knowledge networks. In Jennifer Dy and Andreas Krause (eds.),Proceedings of the 35th International Conference on Machine Learnin...

  77. [85]

    How powerful are graph neural networks?, 2019

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks?, 2019

  78. [86]

    Shedding light on random dropping and oversmoothing

    Han Xuanyuan, Tianxiang Zhao, and Dongsheng Luo. Shedding light on random dropping and oversmoothing. InNeurIPS 2023 Workshop: New Frontiers in Graph Learning, 2023

  79. [87]

    Vishwanathan

    Pinar Yanardag and S.V .N. Vishwanathan. Deep graph kernels. InProceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’15, pp. 1365–1374, New York, NY , USA, 2015. Association for Computing Machinery

  80. [88]

    Hamilton, and Jure Leskovec

    Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L. Hamilton, and Jure Leskovec. Graph convolutional neural networks for web-scale recommender systems. InPro- ceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ’18, ...

  81. [89]

    Graph contrastive learning with augmentations

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (eds.),Advances in Neural Information Processing Systems, volume 33, pp. 5812–5823...

  82. [90]

    Yuning You, Tianlong Chen, Zhangyang Wang, and Yang Shen. When does self-supervision help graph convolutional networks? In Hal Daum ´e III and Aarti Singh (eds.),Proceedings of the 37th International Conference on Machine Learning, volume 119 ofProceedings of Machine Learning ...

  83. [91]

    L2-gcn: Layer-wise and learned efficient training of graph convolutional networks

    Yuning You, Tianlong Chen, Zhangyang Wang, and Yang Shen. L2-gcn: Layer-wise and learned efficient training of graph convolutional networks. In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2124–2132, 2020

  84. [92]

    Bringing your own view: Graph contrastive learning without prefabricated data augmentations

    Yuning You, Tianlong Chen, Zhangyang Wang, and Yang Shen. Bringing your own view: Graph contrastive learning without prefabricated data augmentations. InProceedings of the Fifteenth ACM International Conference on Web Search and Data Mining, WSDM ’22, pp. 1300–1309, New York, ...

  85. [93]

    Pairnorm: Tackling oversmoothing in gnns

    Lingxiao Zhao and Leman Akoglu. Pairnorm: Tackling oversmoothing in gnns. InInterna- tional Conference on Learning Representations, 2020

  86. [94]

    Cold brew: Distilling graph node representations with incomplete or missing neighborhoods

    Wenqing Zheng, Edward W Huang, Nikhil Rao, Sumeet Katariya, Zhangyang Wang, and Karthik Subbian. Cold brew: Distilling graph node representations with incomplete or missing neighborhoods. InInternational Conference on Learning Representations, 2022. 15

  87. [95]

    Towards deeper graph neural networks with differentiable group normalization

    Kaixiong Zhou, Xiao Huang, Yuening Li, Daochen Zha, Rui Chen, and Xia Hu. Towards deeper graph neural networks with differentiable group normalization. InAdvances in neural information processing systems, 2020

  88. [96]

    Dirichlet energy constrained learning for deep graph neural networks.Advances in neural information processing systems, 2021

    Kaixiong Zhou, Xiao Huang, Daochen Zha, Rui Chen, Li Li, Soo-Hyun Choi, and Xia Hu. Dirichlet energy constrained learning for deep graph neural networks.Advances in neural information processing systems, 2021

  89. [97]

    Understanding and resolving performance degradation in deep graph convolutional net- works

    Kuangqi Zhou, Yanfei Dong, Kaixin Wang, Wee Sun Lee, Bryan Hooi, Huan Xu, and Jiashi Feng. Understanding and resolving performance degradation in deep graph convolutional net- works. InProceedings of the 30th ACM International Conference on Information & Knowledge Management, ...

  90. [98]

    Beyond homophily in graph neural networks: Current limitations and effective designs

    Jiong Zhu, Yujun Yan, Lingxiao Zhao, Mark Heimann, Leman Akoglu, and Danai Koutra. Beyond homophily in graph neural networks: Current limitations and effective designs. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (eds.),Advances in Neural Information Proc...

  91. [99]

    Predicting multicellular function through multi-layer tissue networks.Bioinformatics, 33(14):i190–i198, 07 2017

    Marinka Zitnik and Jure Leskovec. Predicting multicellular function through multi-layer tissue networks.Bioinformatics, 33(14):i190–i198, 07 2017. 16 Appendix Table of Contents A Related Works 17 A.1 Methods for Alleviating Over-smoothing . . . . . . . . . . . . . . . . . . . ...

  92. [100]

    These meth- ods successfully addressed the over-smoothing problem, enabling the training of deep GNNs, and performed competitively on several benchmarking datasets

    combines DropEdge and DropMessage together, resulting in a layer-wise sampling scheme that uses a different subgraph for message-aggregation over each feature dimension. These meth- ods successfully addressed the over-smoothing problem, enabling the training of deep GNNs, and ...

  93. [101]

    Residual connections successfully improved the performance of GCN on a range of graph-learning tasks

    proposed the use of residual connections to prevent the smoothing of representations. Residual connections successfully improved the performance of GCN on a range of graph-learning tasks

  94. [102]

    bottleneckedness

    introduced GCN-II, which uses skip connections from the input to all hidden layers. This layer wise propagation rule has allowed for training of ultra-deep networks – up to 64 layers. Some other architectures, like the Jumping Knowledge Network (JKNet) [84] and the Deep Adapti...

  95. [103]

    One concern with graph rewiring methods is that unmoderated densification of the graph, e.g

    proposed the Greedy Total Resistance (GTR) technique, which optimizes the graph’s total re- sistance by greedily adding edges to achieve the greatest improvement. One concern with graph rewiring methods is that unmoderated densification of the graph, e.g. using a fully connect...

  96. [104]

    Flip Markov Chain

    proposed a differentiable graph rewiring layer based on the Lov ´asz bound [55, Corollary 3.3]. Similarly, [6] introduced the Random Local Edge Flip (RLEF) algorithm, which draws inspiration from the “Flip Markov Chain” [22, 56] – a sequence of such steps can convert a connect...

  97. [105]

    Degree c o m p u t a t i o n in Dr op Sen s assumes absence

    showed that the sensitivity between any two nodes in a graph can be bounded using the sum of the powers of the propagation matrix. In Appendix C.1, we extended this bound to random edge- dropping methods with independent edge masks sampled in each layer: EM(1),...,M(L) " ∂z (L...

  98. [106]

    In Table 7, we present Hedges’gstatistic for the statistical tests in Section 5.2

    suggested that an effect size of0.2be considered small,0.5be considered medium, and0.8be considered large. In Table 7, we present Hedges’gstatistic for the statistical tests in Section 5.2. We can clearly see that for homophilic datasets, there is a strongpositive effectof usi...

Pith tools

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