Pith. sign in

REVIEW 4 major objections 5 minor 48 references

DeltaGNN: Graph Neural Network with Information Flow Control

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read One cheap score flags both bottlenecks and heterophilic edges, and DeltaGNN uses it to rewire graphs on the fly.

desk verdict A useful linear-time rewiring heuristic and a solid empirical study, but the theoretical support is a design rationale, not a proof. read the letter →

arxiv 2501.06002 v1 pith:6KS2O333 submitted 2025-01-10 cs.LG

Kevin Mancini , Islem Rekik This is my paper · ORCID
classification cs.LG
keywords graphneuralnetworksover-smoothingover-squashinglong-rangeinteractionsrewiringinformationflownodeclassificationheterophily
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a node-level score computed from the velocity and acceleration of embedding updates during message passing can identify both heterophilic edges and graph bottlenecks at once. This information flow score is built into a rewiring mechanism that filters edges layer by layer with linear overhead, and the resulting DeltaGNN architecture splits the graph into a homophilic part for short-range interactions and a condensed heterophilic part for long-range ones. If correct, standard GNNs could be rewired during training to reduce over-smoothing and over-squashing without attention mechanisms or expensive curvature computations, making deep GNNs practical on large graphs.

What carries the argument

The information flow score $S_u = (V_t[\Delta^2_u] + 1)/(\Delta_u + 1)$ with $m=l=1$, where $\Delta_u$ is the average first delta embedding (the distance between aggregated and transformed neighbor features) and $V_t[\Delta^2_u]$ is the variance over time of the second delta embeddings (the rate of change of the first deltas). The denominator term is meant to flag heterophilic nodes via Lemma 1, the numerator term is meant to flag bottleneck-adjacent nodes via Lemma 2, and Welford's method computes both statistics online so that the per-node overhead stays $O(|V|)$ per layer.

What would settle it

Run a randomly initialized two-layer GCN on a graph built by joining two dense random clusters with a single bridge edge, compute $V_t[\Delta^2_u]$ for every node over several seeds, and check whether the two bridge endpoints consistently fall in the lowest variance quartile; if they do not, the variance term is not a reliable bottleneck signal.

Watch

Extended reading notes

Core claim

The paper's central claim is that the information flow score captures the two failure points of message passing in a single number: a high mean first delta embedding marks a node with low homophily, while a low variance of second delta embeddings marks a node near a bottleneck. DeltaGNN operationalizes this by removing low-score edges during message passing, then condensing the high-score nodes into a small heterophilic graph to preserve long-range interactions, and finally combining both homophilic and heterophilic aggregations before classification.

Load-bearing premise

The central assumption is that a node's proximity to a bottleneck reliably shows up as low variance in its second-order embedding updates, so that the variance term really tracks connectivity rather than just feature noise.

Editorial extensions

If this is right

  • Any GNN can insert information flow control as a plug-in edge-filtering module, gaining bottleneck and heterophily detection at linear cost per layer.
  • Deep GNNs can be trained on graphs where attention-based transformers run out of memory and where curvature-based rewiring runs out of time.
  • The dual homophilic and heterophilic aggregation lets a single model perform well on graphs with both high and low homophilic ratios, where standard GCN and GAT degrade.
  • Removing low-score edges while hill-climbing the mean score preserves graph sparsity, so the rewiring does not densify the graph.
  • Because the information flow score is computed during message passing, it has zero preprocessing overhead, unlike centrality and curvature measures.

Reading between the lines

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

  • If the variance-connectivity correspondence behind Lemma 2 holds across diverse topologies, the same score could be reused for graph pooling or explainability by ranking nodes rather than edges.
  • The score's dependence on the feature distance function $d$ may make it sensitive to feature scaling; normalizing features or using cosine distance could change which edges are flagged, a testable extension.
  • The heterophilic graph condensation step effectively performs a learned, score-based sampling of a fully connected subgraph, suggesting that explicit long-range coupling can be replaced by a small complete graph over informative nodes.
  • The 'one-for-all' claim is tested on node classification; edge-level and graph-level benchmarks would need to confirm that the linear-cost advantage transfers.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes information flow control (IFC), a graph-rewiring mechanism for GNNs, guided by a per-node information flow score (IFS) defined as S_u = (m * V_t[Δ²_u] + 1) / (l * Δ_u + 1), where Δ_u is the time-averaged first delta embedding and V_t[Δ²_u] is the variance of second delta embeddings. The authors claim that low IFS identifies nodes adjacent to bottlenecks and heterophilic edges, and they integrate IFC into a new architecture, DeltaGNN, with dual homophilic and heterophilic aggregation. The paper reports accuracy, specificity, and complexity comparisons on ten datasets, including large and dense graphs, and argues that IFS has O(|V|) overhead, making it more scalable than curvature- or centrality-based rewiring. The central theoretical support is provided by Lemma 1 (high average Δ implies low homophily) and Lemma 2 (low variance of Δ² implies bottleneck adjacency).

Significance. If the theoretical claims were valid, the paper would contribute a novel, scalable, embedding-aware connectivity measure that jointly targets over-smoothing and over-squashing, and DeltaGNN's strong empirical results on diverse and dense graphs would be a useful practical advance. The authors are to be credited for a comprehensive experimental setup: ten datasets with varying homophily and density, careful grid-search tuning with fixed layer counts, multiple rewiring baselines, reported confidence intervals, public code, and resource comparisons. The idea of using temporal velocity/acceleration of embeddings as a rewiring signal is original and worth pursuing. However, the two lemmas that ground the IFS are not rigorously established; the bottleneck-detection half rests entirely on an asserted variance-connectivity correspondence. Thus the theoretical contribution, as written, does not support the paper's claims, although the empirical contribution remains valuable.

major comments (4)
  1. [Section IV-A, Lemma 2] The proof of Lemma 2 does not derive the claimed implication 'low connectivity ⇒ low V_t[Δ²_u]'. After citing [14], it states: 'This intuitive observation implies that V_t[Δ²_v] must necessarily be greater than V_t[Δ²_u].' No argument is given connecting the temporal variance of the second delta embeddings to the message-passing dynamics, spectral properties, nonlinearities, or the fact that the graph is rewired at every layer in Algorithm 1. Because the Δ² sequence is generated on a time-varying graph, convergence results for fixed diffusion processes do not apply. If this variance-connectivity correspondence fails, the IFS in Eq. (4) will not systematically rank bottleneck-adjacent nodes low, and the bottleneck-detection half of the proposed rewiring loses its stated basis. The lemma needs either a rigorous derivation under explicit assumptions or a downgrade to a heuristic claim.
  2. [Section IV-A, Lemma 1] The proof of Lemma 1 has several load-bearing defects. (a) The threshold ρ is chosen as ρ > sup_{h∈[H,1]} U(h)_u, but U(h)_u depends on the specific node u through its neighborhood and degree; the lemma asserts a single positive lower-bound ρ valid for all nodes, which is not established. (b) The definition of the set S on page 3 contains a malformed condition: 'm ∈ s(v)' treats the feature vector s(v) as a set, and the intended constraint is unclear. (c) The final deduction has the wrong direction: the text shows that Δ_t^u > ρ implies H_u < H, and then writes 'Δ_t^u > ρ → Σ_i Δ_i^u > Tρ → Δ_u > ρ', but the lemma requires the converse — that Δ_u > ρ implies H_u < H. The implication from a single-layer bound to the time-averaged bound is not proved. These issues invalidate the homophily-detection part of the IFS.
  3. [Section IV-A, paragraph after Lemma 2; Section VI-C] The manuscript itself acknowledges a significant limitation: when nodes belong to very different neighborhoods, the IFS 'will not deterministically find the optimal solution and its accuracy will be negatively affected by the amount of noise introduced.' This is not reflected in the abstract's claim that DeltaGNN is 'the first scalable and generalizable approach for detecting long-range and short-range interactions.' The numerical experiment in Section VI-C is a single 14-node toy graph; Table IV reports homophily improvements on CORA but does not directly measure whether the removed edges are actually at bottlenecks. Thus the empirical results do not substitute for the missing derivation of the variance-bottleneck correspondence, and the stated limitation should qualify the generalizability claim.
  4. [Section V-B and Table III] The claimed O(|V|) time complexity for the IFS is not fully justified. While computing the per-node scores may be O(|V|·d_V·T), Algorithm 1 also removes the K(t,θ) edges with the lowest scores at each layer, which requires a selection or sorting step of at least O(|V| log |V|) in general, and updating the adjacency matrix can touch O(|E|) entries. The empirical timings in Table VIII are useful, but the theoretical 'O(|V|) overhead' claim in Section V-B is stronger than what is demonstrated.
minor comments (5)
  1. [Section II-D and Algorithm 1] The text after Eq. (4) states that the mean and variance are computed using an exponential moving average to reduce noise, but Algorithm 1 (lines 13–14) uses Welford's method for the arithmetic mean and variance. These are different update rules; please clarify which is used in the experiments.
  2. [Definition IV.1] The aggregation symbol in the definition of Δ_t^u appears garbled in the text ('d( M v∈N (u) Mt_v, Mt_u)'); the intended formula with ⊕ should be written explicitly. Similarly, in the proof of Lemma 1 the symbol 'L' appears in place of the aggregation operator.
  3. [Table VI] For CiteSeer, the training set size is listed as 18,217, which exceeds the total number of nodes (3,327). This is clearly a typo and should be corrected to the actual split used.
  4. [Table VIII] For GCN + rewiring(EC) on CiteSeer, the epoch time is 0.0014, identical to the GCN baseline; since rewiring is expected to add overhead, this likely indicates a transcription error.
  5. [Section VI-A] The phrase 'the IFS is a one-for-all connectivity measure' is an overstatement; the experiments show that IFS performs well on several datasets, but 'one-for-all' is not established, especially given the acknowledged sensitivity to neighborhood differences.

Circularity Check

2 steps flagged · score 4.0 of 10

The bottleneck-detection half of the IFS rests on Lemma 2, whose proof assumes the conclusion; Lemma 1's threshold is defined to force its own implication, but the empirical benchmark against external baselines is independent.

  1. other [Section IV-A, Lemma 2 (Proof)]
    "Now, we can distinguish the two cases: v will converge faster, leading to high values of (∆2 v)t for small values of t and low values later in time. On the other hand, for node u, the values of (∆2 u)t will vary more slowly over time due to its slower convergence. This intuitive observation implies that Vt[∆2 v] must necessarily be greater than Vt[∆2 u]."

    Lemma 2 is supposed to establish that bottleneck-adjacent nodes have low variance of the second delta embeddings, which is then used to justify the numerator of the information flow score. The proof, however, asserts this property as an 'intuitive observation' rather than deriving it from the message-passing equations, the graph spectrum, or the cited [14] result. The sentence 'nodes near bottlenecks tend to have constrained communication paths ... and are consequently less likely to experience significant fluctuations in their embedding values' is exactly the variance-to-bottleneck correspondence that the lemma must prove, and the final 'must necessarily' is a restatement of the lemma's conclusion.

  2. self definitional [Section IV-A, Lemma 1 (Proof)]
    "Now, for any homophilic ratio H ∈[0, 1] we can choose a ρ > suph∈[H,1] U (hu)u such that any node u with ∆t u > ρ will have Hu < H."

    The threshold ρ is constructed after defining U(h_u)_u as the supremum of the first delta embeddings over all feature assignments that respect homophily h. Choosing ρ above sup_{h∈[H,1]} U(h_u)_u makes the implication 'Δ > ρ ⇒ H_u < H' true by construction; the lemma does not derive a relationship between Δ and homophily from the dynamics of message passing. It is a tautological existence statement. The later inference that 'nodes with the highest values of Δ will likely have the lowest homophilic rates' therefore rests on a definitional threshold, not on an independent property of the delta embeddings.

full rationale

The paper's central theoretical derivation chain is: Lemma 1 maps high Δ to low homophily, Lemma 2 maps low V_t[Δ²] to bottleneck adjacency, and Equation 4 combines these into the information flow score. Lemma 2 is the load-bearing step for the claim that the IFS identifies bottlenecks, and its proof is circular: the 'intuitive observation' that bottleneck-adjacent nodes experience less fluctuation in their embedding values is precisely the variance-bottleneck correspondence that the lemma must prove, and the final 'must necessarily' is asserted rather than derived. Lemma 1 is also definitional: ρ is chosen as a bound on the suprema U(h_u)_u, so the implication is true by construction of the threshold. Neither step is a fit of data, and there is no load-bearing self-citation chain: [14] is an external curvature paper and [30] is only cited as a predecessor architecture. The empirical sections (Tables I, II, IV, V, Figures 5-7) compare against external baselines and show that the IFS improves homophily and accuracy on benchmark datasets, which is independent support regardless of the lemma defects. Thus the circularity is partial: the theoretical justification for the score's bottleneck semantics is circular/tautological, but the method's empirical evaluation is self-contained and the central contribution does not reduce entirely to its inputs.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No physical entities are introduced. The free parameters are the two hand-set multipliers m and l plus the hill-ascent parameter theta. The central claim rests on five stated assumptions, two of which are effectively ad hoc to the paper's own desired conclusion.

free parameters (3)
  • m = 1 (set by hand)
    Multiplier on the variance term in the information flow score (Eq. 4). Set to 1 for equal priority between bottleneck detection and heterophily detection, not fitted to data.
  • l = 1 (set by hand)
    Multiplier on the mean delta term in the information flow score (Eq. 4). Set to 1 for equal priority, not fitted to data.
  • theta = optimized online via hill ascent
    Controls the fraction of edges removed at each layer through K(t, theta). Adjusted by hill ascent on the mean node score (Algorithm 1), so it is a fitted parameter of the filtering mechanism.
assumptions (5)
  • domain assumption There exists a unique function phi: M -> C assigning each possible feature vector a label (Lemma 1).
    Needed to define the constraint set S and threshold rho; not guaranteed for real learning problems where labels are only observed on graph nodes.
  • domain assumption Node embeddings lie on a compact manifold M with a continuous distance d (Lemma 1, Hopf-Rinow).
    Compactness justifies the Extreme Value Theorem step; not established for trained GNN embeddings.
  • domain assumption The first delta sequence Delta^t_u converges to zero for all nodes as t grows (Lemma 1).
    This is essentially the over-smoothing limit the method aims to control, yet it is used at the start of the proof.
  • ad hoc to paper There exists mu such that c(u) < mu iff u is adjacent to an edge bottleneck (Lemma 2).
    This assumption encodes the desired conclusion of the lemma; no construction of mu is given.
  • domain assumption High-score nodes have reliable representations and can seed the heterophilic condensation graph (Section V-A).
    The paper states that high-score nodes are likely to have reliable representations, but provides no proof; the long-range interaction recovery depends on this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DeltaGNN: Graph Neural Network with Information Flow Control." pith.science (2026). https://pith.science/paper/6KS2O333

@misc{pith2026250106002,
  author       = {Pith},
  title        = {Pith review of: DeltaGNN: Graph Neural Network with Information Flow Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6KS2O333}},
  note         = {Machine review of arXiv:2501.06002}
}
read the original abstract

Graph Neural Networks (GNNs) are popular deep learning models designed to process graph-structured data through recursive neighborhood aggregations in the message passing process. When applied to semi-supervised node classification, the message-passing enables GNNs to understand short-range spatial interactions, but also causes them to suffer from over-smoothing and over-squashing. These challenges hinder model expressiveness and prevent the use of deeper models to capture long-range node interactions (LRIs) within the graph. Popular solutions for LRIs detection are either too expensive to process large graphs due to high time complexity or fail to generalize across diverse graph structures. To address these limitations, we propose a mechanism called \emph{information flow control}, which leverages a novel connectivity measure, called \emph{information flow score}, to address over-smoothing and over-squashing with linear computational overhead, supported by theoretical evidence. Finally, to prove the efficacy of our methodology we design DeltaGNN, the first scalable and generalizable approach for detecting long-range and short-range interactions. We benchmark our model across 10 real-world datasets, including graphs with varying sizes, topologies, densities, and homophilic ratios, showing superior performance with limited computational complexity. The implementation of the proposed methods are publicly available at https://github.com/basiralab/DeltaGNN.

Figures

Figures reproduced from arXiv: 2501.06002 by the authors.

Figure 1
Figure 1. Illustration of a graph with a bottleneck. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of a GNN model with information flow control. The figure illustrates how the information flow control mechanism integrates with a standard GNN to filter the graph and increase the mean node score. Simultaneously, the GNN learns to disentangle the node features, ensuring that the output embeddings can be easily classified using a readout layer. The novel components are highlighted in violet. Notably, the mea… view at source ↗
Figure 3
Figure 3. The DeltaGNN pipeline consists of: (a) a [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Validation accuracy (red) and specificity (blue) convergence during training epochs for DeltaGNN [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Illustrations of first delta embeddings, second delta embeddings, and information flow score on a small graph with bottlenecks and heterophilic edges. We set m = 1 and l = 1 to detect heterophilic bottlenecks, using the Euclidean distance as the distance metric d. The …
Figure 6
Figure 6. Figure 6: Filtered graph with updated information flow score. After removing bottleneck edges and heterophilic node interactions, the mean node score increased significantly, demonstrating the effectiveness of the edge filtering [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Homophilic ratio density distribution shifts during topological edge-filtering and heterophilic graph condensation using seven distinct connectivity [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Comparison of graph-rewiring techniques for alleviating over-smoothing and over-squashing. Distinct node colors represent different node classes. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 31 canonical work pages

  1. [14]

    Revisiting over-smoothing and over- squashing using ollivier-ricci curvature,

    K. Nguyen, N. M. Hieu, V . D. Nguyen, N. Ho, S. Osher, and T. M. Nguyen, “Revisiting over-smoothing and over- squashing using ollivier-ricci curvature,” in International Conference on Machine Learning . PMLR, 2023, pp. 25 956–25 979

    work page 2023

  2. [1]

    The graph neural network model,

    F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfardini, “The graph neural network model,” IEEE transactions on neural networks , vol. 20, no. 1, pp. 61– 80, 2008

    work page 2008

  3. [2]

    Semi-supervised classifica- tion with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classifica- tion with graph convolutional networks,” Sep 09, 2016. [Online]. Available: https://arxiv.org/abs/1609.02907

    arXiv 2016

  4. [3]

    A comprehensive survey on graph neural networks,

    Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and P. S. Yu, “A comprehensive survey on graph neural networks,” CoRR, vol. abs/1901.00596, 2019. [Online]. Available: http://arxiv.org/abs/1901.00596

    arXiv 1901

  5. [4]

    Hierarchical graph convolutional networks for semi-supervised node classification,

    F. Hu, Y . Zhu, S. Wu, L. Wang, and T. Tan, “Hierarchical graph convolutional networks for semi-supervised node classification,” arXiv preprint arXiv:1902.06667 , 2019

    arXiv 1902

  6. [5]

    Message passing neural networks,

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, “Message passing neural networks,” p. 199, 2020

    work page 2020

  7. [6]

    Predicting multicellular function through multi-layer tissue networks

    M. Zitnik and J. Leskovec, “Predicting multicellular function through multi-layer tissue networks,” CoRR, vol. abs/1707.04638, 2017. [Online]. Available: http: //arxiv.org/abs/1707.04638

    work page Pith review arXiv 2017

  8. [7]

    Graph-based deep learning for medical diagnosis and analysis: Past, present and future,

    D. Ahmedt-Aristizabal, M. A. Armin, S. Denman, C. Fookes, and L. Petersson, “Graph-based deep learning for medical diagnosis and analysis: Past, present and future,” Sensors, vol. 21, no. 14, p. 4758, jul 2021. [Online]. Available: https://doi.org/10.3390% 2Fs21144758

    work page 2021

Show all 48 references
  1. [8]

    Graph neural networks in network neuroscience,

    A. Bessadok, M. A. Mahjoub, and I. Rekik, “Graph neural networks in network neuroscience,” 1 2015

    2015

  2. [9]

    Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classifica- tion,

    J. Yang, R. Shi, D. Wei, Z. Liu, L. Zhao, B. Ke, H. Pfis- ter, and B. Ni, “Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classifica- tion,” Scientific Data, vol. 10, no. 1, p. 41, 2023

    2023

  3. [10]

    Deeper insights into graph convolutional networks for semi-supervised learning,

    Q. Li, Z. Han, and X.-M. Wu, “Deeper insights into graph convolutional networks for semi-supervised learning,” in Proceedings of the AAAI conference on artificial intelligence, vol. 32, no. 1, 2018

    2018

  4. [11]

    On the bottleneck of graph neural networks and its practical implications,

    U. Alon and E. Yahav, “On the bottleneck of graph neural networks and its practical implications,” arXiv preprint arXiv:2006.05205, 2020

    2006 arXiv

  5. [12]

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

    Z. Wu, P. Jain, M. Wright, A. Mirhoseini, J. E. Gonza- lez, and I. Stoica, “Representing long-range context for graph neural networks with global attention,” Advances in Neural Information Processing Systems , vol. 34, pp. 13 266–13 279, 2021

    2021

  6. [13]

    A gnn architecture with local and global-attention fea- ture for image classification,

    Z. Fei, J. Guo, H. Gong, L. Ye, E. Attahi, and B. Huang, “A gnn architecture with local and global-attention fea- ture for image classification,” IEEE Access, 2023

    2023

  7. [15]

    Locality-aware graph-rewiring in gnns,

    F. Barbero, A. Velingker, A. Saberi, M. Bronstein, and F. Di Giovanni, “Locality-aware graph-rewiring in gnns,” arXiv preprint arXiv:2310.01668 , 2023

    2023 arXiv

  8. [16]

    A sur- vey on oversmoothing in graph neural networks,

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

    2023 arXiv

  9. [17]

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

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

    2020

  10. [18]

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

    F. Di Giovanni, L. Giusti, F. Barbero, G. Luise, P. Lio, and M. M. Bronstein, “On over-squashing in message passing neural networks: The impact of width, depth, and topology,” in International Conference on Machine Learning. PMLR, 2023, pp. 7865–7885

    2023

  11. [19]

    Ricci curvature of markov chains on metric spaces,

    Y . Ollivier, “Ricci curvature of markov chains on metric spaces,” Journal of Functional Analysis , vol. 256, no. 3, pp. 810–864, 2009

    2009

  12. [20]

    Ollivier-ricci curvature-based method to community detection in com- plex networks,

    J. Sia, E. Jonckheere, and P. Bogdan, “Ollivier-ricci curvature-based method to community detection in com- plex networks,” Scientific reports, vol. 9, no. 1, p. 9800, 2019

    2019

  13. [21]

    Note on a method for calculating cor- rected sums of squares and products,

    B. P. Welford, “Note on a method for calculating cor- rected sums of squares and products,” Technometrics, vol. 4, no. 3, pp. 419–420, 1962

    1962

  14. [22]

    Graph transformer networks,

    S. Yun, M. Jeong, R. Kim, J. Kang, and H. J. Kim, “Graph transformer networks,” Advances in neural in- formation processing systems , vol. 32, 2019

    2019

  15. [23]

    A generalization of transformer networks to graphs,

    V . P. Dwivedi and X. Bresson, “A generalization of transformer networks to graphs,” arXiv preprint arXiv:2012.09699, 2020

    2012 arXiv

  16. [24]

    Under- standing oversquashing in gnns through the lens of effec- tive resistance,

    M. Black, Z. Wan, A. Nayyeri, and Y . Wang, “Under- standing oversquashing in gnns through the lens of effec- tive resistance,” in International Conference on Machine Learning. PMLR, 2023, pp. 2528–2547

    2023

  17. [25]

    Diffwire: Inductive graph rewiring via the lov\’asz bound,

    A. Arnaiz-Rodr ´ıguez, A. Begga, F. Escolano, and N. Oliver, “Diffwire: Inductive graph rewiring via the lov\’asz bound,” arXiv preprint arXiv:2206.07369, 2022

    2022 arXiv

  18. [26]

    Fosr: First-order spectral rewiring for addressing oversquash- ing in gnns,

    K. Karhadkar, P. K. Banerjee, and G. Mont ´ufar, “Fosr: First-order spectral rewiring for addressing oversquash- ing in gnns,” arXiv preprint arXiv:2210.11790 , 2022

    2022 arXiv

  19. [27]

    Two sides of the same coin: Heterophily and oversmoothing in graph convolutional neural networks,

    Y . Yan, M. Hashemi, K. Swersky, Y . Yang, and D. Koutra, “Two sides of the same coin: Heterophily and oversmoothing in graph convolutional neural networks,” in 2022 IEEE International Conference on Data Mining (ICDM). IEEE, 2022, pp. 1287–1292

    2022

  20. [28]

    Ueber den begriff der vollst¨andigen differentialgeometrischen fl ¨ache,

    H. Hopf and W. Rinow, “Ueber den begriff der vollst¨andigen differentialgeometrischen fl ¨ache,” Com- mentarii Mathematici Helvetici , vol. 3, no. 1, pp. 209– 18 225, 1931

    1931

  21. [29]

    Bolzano and uniform continuity,

    P. Rusnock and A. Kerr-Lawson, “Bolzano and uniform continuity,” Historia Mathematica , vol. 32, no. 3, pp. 303–311, 2005. [Online]. Available: https://www.sciencedirect.com/science/article/ pii/S0315086004000849

    2005

  22. [30]

    Duognn: Topology- aware graph neural network with homophily and heterophily interaction-decoupling,

    K. Mancini and I. Rekik, “Duognn: Topology- aware graph neural network with homophily and heterophily interaction-decoupling,” arXiv preprint arXiv:2409.19616, 2024

    2024 arXiv

  23. [31]

    Representation learning on graphs with jumping knowledge networks,

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

    2018

  24. [32]

    How powerful are graph neural networks?

    K. Xu, W. Hu, J. Leskovec, and S. Jegelka, “How powerful are graph neural networks?” arXiv preprint arXiv:1810.00826, 2018

    2018 arXiv

  25. [33]

    Graph attention networks,

    P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Lio, Y . Bengioet al., “Graph attention networks,” stat, vol. 1050, no. 20, pp. 10–48 550, 2017

    2017

  26. [34]

    Understanding over-squashing and bottlenecks on graphs via curvature,

    J. Topping, F. Di Giovanni, B. P. Chamberlain, X. Dong, and M. M. Bronstein, “Understanding over-squashing and bottlenecks on graphs via curvature,” arXiv preprint arXiv:2111.14522, 2021

    2021 arXiv

  27. [35]

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

    J. Zhu, Y . Yan, L. Zhao, M. Heimann, L. Akoglu, and D. Koutra, “Beyond homophily in graph neural networks: Current limitations and effective designs,” Advances in neural information processing systems, vol. 33, pp. 7793– 7804, 2020

    2020

  28. [36]

    Geom-gcn: Geometric graph convolutional networks,

    H. Pei, B. Wei, K. C.-C. Chang, Y . Lei, and B. Yang, “Geom-gcn: Geometric graph convolutional networks,” arXiv preprint arXiv:2002.05287 , 2020

    2002 arXiv

  29. [37]

    Masked label prediction: Unified message passing model for semi-supervised classification,

    Y . Shi, Z. Huang, S. Feng, H. Zhong, W. Wang, and Y . Sun, “Masked label prediction: Unified message passing model for semi-supervised classification,” arXiv preprint arXiv:2009.03509, 2020

    2009 arXiv

  30. [38]

    Nagphormer: A tokenized graph transformer for node classification in large graphs,

    J. Chen, K. Gao, G. Li, and K. He, “Nagphormer: A tokenized graph transformer for node classification in large graphs,” arXiv preprint arXiv:2206.04910 , 2022

    2022 arXiv

  31. [39]

    Automating the construction of internet portals with machine learning,

    A. K. McCallum, K. Nigam, J. Rennie, and K. Seymore, “Automating the construction of internet portals with machine learning,” Information Retrieval , vol. 3, pp. 127–163, 2000

    2000

  32. [40]

    Revisiting semi-supervised learning with graph embeddings,

    Z. Yang, W. Cohen, and R. Salakhudinov, “Revisiting semi-supervised learning with graph embeddings,” in International conference on machine learning . PMLR, 2016, pp. 40–48

    2016

  33. [41]

    Affordable graph neural net- work framework using topological graph contraction,

    C. Adnel and I. Rekik, “Affordable graph neural net- work framework using topological graph contraction,” in Workshop on Medical Image Learning with Limited and Noisy Data. Springer, 2023, pp. 35–46

    2023

  34. [42]

    Falcon: Feature-label constrained graph net col- lapse for memory-efficient gnns,

    ——, “Falcon: Feature-label constrained graph net col- lapse for memory-efficient gnns,” IEEE Transactions on Neural Networks and Learning Systems , 2024

    2024

  35. [43]

    Deep learning,

    Y . LeCun, Y . Bengio, and G. Hinton, “Deep learning,” nature, vol. 521, no. 7553, pp. 436–444, 2015

    2015

  36. [44]

    Welford’s algorithm for weighted statistics,

    A. A. Efanov, S. A. Ivliev, and A. G. Shagraev, “Welford’s algorithm for weighted statistics,” in 2021 3rd International Youth Conference on Radio Electronics, Electrical and Power Engineering (REEPE) . IEEE, 2021, pp. 1–5

    2021

  37. [45]

    Complex network metrics: Can deep learning keep up with tailor-made reference algorithms?

    S. Wandelt, X. Shi, and X. Sun, “Complex network metrics: Can deep learning keep up with tailor-made reference algorithms?” IEEE Access, vol. 8, pp. 68 114– 68 123, 2020

    2020

  38. [46]

    Demystifying structural disparity in graph neural networks: Can one size fit all?

    H. Mao, Z. Chen, W. Jin, H. Han, Y . Ma, T. Zhao, N. Shah, and J. Tang, “Demystifying structural disparity in graph neural networks: Can one size fit all?” Advances in neural information processing systems , vol. 36, 2024

    2024

  39. [47]

    Automatic differentiation in pytorch,

    A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer, “Automatic differentiation in pytorch,” 2017

    2017

  40. [48]

    Fast graph representa- tion learning with pytorch geometric,

    M. Fey and J. E. Lenssen, “Fast graph representa- tion learning with pytorch geometric,” arXiv preprint arXiv:1903.02428, 2019. Kevin Mancini received his B.S. degree in Com- puter Engineering from the Alma Mater Studiorum – University of Bologna, Italy, in 2018, and his M.S. ...

    2025 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.