A Sheaf-Theoretic and Topological Perspective on Complex Network Modeling and Attention Mechanisms in Graph Neural Models

Chuan-Shen Hu

REVIEW 4 major objections 4 minor 1 cited by

A Sheaf-Theoretic and Topological Perspective on Complex Network Modeling and Attention Mechanisms in Graph Neural Models

T0 review · 4 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Trained GAT attention weights can be read as a cellular sheaf, giving a topological model of feature diffusion.

desk verdict A short, correct, mostly definitional paper that translates GAT attention weights into a cellular sheaf and defines a harmonic-set filtration; the claimed training-dynamics diagnostics are not demonstrated. read the letter →

arxiv 2601.21207 v4 pith:QHCTWWBM submitted 2026-01-29 cs.LG cs.AImath.AT

Chuan-Shen Hu This is my paper · ORCID

classification cs.LGcs.AImath.AT MSC 18F2055N3168T07

keywords cellularsheavesgraphattentionnetworkssheafLaplacianharmonicsetstopologicaldataanalysispersistencebarcodesfeaturediffusiongeometricdeeplearning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that the trained attention weights of a graph attention network, together with its node features, form a cellular sheaf—an object that assigns vector spaces to nodes and edges and linear maps to their incidence relations. On that basis it defines 'harmonic' edges and nodes where transferred features agree, and shows these form a subgraph that equals the whole graph exactly when the feature vector is a global section. It then introduces a multiscale, TDA-style filtration of harmonic substructures as a tool for quantifying local-to-global signal alignment. The payoff would be a topological vocabulary for feature diffusion in graph neural models, with potential diagnostics for oversmoothing and heterophily.

What carries the argument

The cellular sheaf is the central object: a functor from the graph viewed as a poset (with face relations v⊴e) to finite-dimensional real vector spaces, assigning a stalk R^d to every node and edge and a restriction linear map along each incidence. In the GAT construction the restriction maps are scalar multiplications by attention weights. The sheaf coboundary and sheaf Laplacian then carry the message-passing and global-section structure; harmonic sets are defined by zero entries of the coboundary, and the epsilon-filtration is level-set filtration of the sheaf norm.

What would settle it

Train a GAT on a real graph dataset, extract the sheaf and its epsilon-filtration at a fixed checkpoint, and compute the persistence barcode. Then randomly permute the attention weights among edges of similar degree and recompute the barcode; if the barcode (or the resulting harmonic subgraphs) does not change in a way that tracks the model's classification behavior or over-smoothing, the claimed diagnostic value of the sheaf structure would not hold.

Watch

Extended reading notes

Core claim

Theorem 1: Any GAT triple (G, (s_vi), W), with d-dimensional node features and attention weights w_ij, defines a cellular sheaf F on G with F_v = F_e = R^d and restriction map F_{vi⊴{vi,vj}} equal to scalar multiplication by w_ij. Consequently the GAT aggregation rule is exactly the sheaf coboundary/diffusion operation. Definition 1 and Theorem 2 characterize local consistency: an edge is harmonic when the two endpoint features are carried to the same edge stalk; the harmonic set is a subgraph that is all of G iff the signal is a global section. Theorem 3 and Corollary 2 build an epsilon-filtration of harmonic sets whose persistence barcode can be computed.

Load-bearing premise

The framework treats the trained attention weights as fixed, signal-independent scalar restriction maps, even though in a real GAT those weights are computed from the node features via softmax, so the sheaf is not a fixed structure during training.

Editorial extensions

If this is right

Every trained GAT snapshot carries a sheaf Laplacian whose zero eigenspace is the space of global sections—features perfectly consistent with the attention weights.
The harmonic subgraph Har(s) equals the whole graph precisely when the signal is a global section, connecting local agreement to global consistency.
The epsilon-filtration Har_epsilon(s) yields a persistence barcode, giving a one-dimensional topological summary of how feature alignment appears across scales.
The construction extends beyond scalar weights: the same formalism supports matrix-valued restriction maps, recovering more general sheaf-based attention aggregations.
Because the framework tracks local alignment only, it is applicable to any graph neural model, not just GATs, whenever edge weights are available.

Reading between the lines

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

Inference: The sheaf construction is most defensible as a snapshot description of a frozen trained model; during training the attention weights are functions of the features, so the restriction maps are signal-dependent and the sheaf is not fixed. Treating it as a static sheaf for training-dynamics analysis would require re-deriving the construction on a co-evolving family.
Inference: The harmonic-set filtration is formally a level-set filtration of the vector-valued sheaf coboundary, so existing TDA stability results for filtrations may apply, suggesting the barcode is stable under small perturbations of features and weights.
Inference: A testable extension would train GATs on standard node-classification datasets, compute persistence barcodes of the harmonic filtration at checkpoints, and check whether barcode statistics (e.g., total persistence) track oversmoothing or accuracy—something the paper proposes as future work but does not run.

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.

Desk Editor's Note

read the letter

Colleague — you asked what I make of this one. Short version: it's a clean but thin theory note. The three theorems are correct, and the paper is honest about what it doesn't do. If your interest is sheaf-theoretic formalizations of GNN attention, you'll get something; if you're expecting new empirical diagnostics, you won't.

What's actually new: Theorem 1 really does construct a cellular sheaf from a GAT triple by setting restriction maps to scalar attention weights. That's a genuinely useful reformulation, even if it's immediate once you see it — it makes GAT message passing a special case of sheaf-theoretic diffusion, and it's nice to have that spelled out. The harmonic set definition and the filtration from Definition 2 and Theorem 3 are also clean, and the proofs are airtight. The paper reads like a competent mathematician's worked example, not an obfuscated cluster.

The soft spots: First, the 'analysis of training dynamics' claim in the abstract and introduction is overreach. The framework treats the sheaf as fixed, but in a real GAT the attention weights are functions of the node features, so the sheaf co-evolves with the signal during training. The paper itself concedes this in the Future Work section, but that caveat sits far from the strong introductory language. Second, the filtration is just sublevel sets of the coboundary norm — a standard TDA move — and no experiment, even a toy one, shows that the persistence barcode tracks oversmoothing, heterophily, or any actual GNN behavior. The diagnostic value is asserted, not established. Third, labels like 'opens several promising directions' lean on the aspirational; the contribution is a vocabulary and a construction, not a method with demonstrated payoff.

That said, I don't see a load-bearing flaw. The mathematics is correct, the citations to sheaf neural networks and TDA are appropriate, and the author doesn't hide the lack of validation. It's a conditional accept at best, but the condition is real: someone needs to instantiate this on trained GATs and compare the barcodes to metrics like oversmoothing. Until then it's a conceptual tool, not a diagnostic.

For peer review: yes, I'd send this to a competent referee. The subfield benefits from clear formal translations, and a good referee can push the author to either temper the claims or add a real experiment. I wouldn't reject it on sight.

Net: cite it if you work on sheaf-based GNN theory; skip it if you want empirical tools. I'd bring it to the reading group if someone else already read the sheaf literature, but I wouldn't assign it as a standalone introduction.

Referee Report

4 major / 4 minor

Summary. The paper introduces a cellular-sheaf model for graph attention networks. For a fixed GAT triple (G, (s_{v_i}), W), Theorem 1 assigns stalk R^d to every vertex and edge and takes the restriction map for each incidence to be scalar multiplication by the corresponding attention weight. The paper then defines harmonic edges/nodes as those whose sheaf-coboundary residual is zero, and epsilon-harmonic sets by thresholding residual norms. Theorem 2 links the full harmonic set to global sections, and Theorem 3/Corollary 2 show that increasing epsilon yields a filtration of subgraphs that can be summarized by persistent homology. The manuscript contains no experiments; its stated aim is to enable future analysis of training dynamics and provide a TDA-based diagnostic.

Significance. The observation that GAT attention weights themselves form a cellular sheaf is a clean dictionary between GNN attention and sheaf theory, and the harmonic-set filtration is a natural sublevel-set construction. However, the mathematical content is largely definitional: Theorem 1 is a construction, Theorem 2 restates the definition of a global section, and Theorem 3 is monotonicity of sublevel sets. The potentially novel part is the application to training dynamics, but this is not developed. No experiments, algorithms, or falsifiable predictions are provided. With its claims appropriately narrowed, the paper could serve as a short theory note; in its current form it overclaims a 'joint characterization of GDL and TDL architectures' and 'insights' into oversmoothing and heterophily without support.

major comments (4)

[Theorem 1 (Eq. (6))] The construction of the cellular sheaf is underspecified for undirected edges. Since W=(w_ij) is not symmetric, an edge {v_i,v_j} requires two restriction maps, one from each endpoint, and these must be scalar multiplication by w_ij and w_ji respectively. The statement currently defines only F_{v_i⊴{v_i,v_j}} via w_ij, leaving the other incidence undefined and the notation dependent on an arbitrary ordering. Fix by fixing a total order on V and writing F_{v_i⊴{v_i,v_j}}=w_ij I and F_{v_j⊴{v_i,v_j}}=w_ji I.
[Theorem 1 / Definition 1] The sheaf in Theorem 1 is signal-dependent: in GAT, each w_ij is computed from the node features via softmax(LeakyReLU(a^T[W h_i || W h_j])), so the restriction maps are functions of the very cochain s whose harmonicity is being measured. Theorem 3 therefore applies to a fixed sheaf F only; it does not describe the training trajectory in which s and W co-evolve. The paper's motivation—'distribution and diffusion behavior ... during training'—is thus not addressed; indeed the Future Work section states that the relation to learning behavior 'remains to be explored.' This gap is load-bearing for the claimed contribution.
[Corollary 2 / Future Work] The claimed diagnostic value of the harmonic filtration is not demonstrated. No experiments are reported; no algorithm or complexity analysis for computing the filtration or persistence barcode is given; and no connection is made to oversmoothing, heterophily, or node classification. The paper only asserts that barcodes 'can be computed.' At minimum, a synthetic experiment on a GAT-trained model (or a precise conjecture with a toy example) is needed to show that Har^epsilon(s) carries information beyond the definition.
[Abstract / Main Results] The abstract's claim that the framework 'enables a joint characterization of GDL and TDL architectures' is not supported by the body. The paper treats only graph-level GAT; simplicial and cell complexes are deferred to future work. Either add the promised higher-dimensional analysis or substantially weaken this claim.

minor comments (4)

[Mathematical Background] The direction of restriction maps is inconsistent: the Definition says F_{v,e}: F_e -> F_v, while Equation (1) and Theorem 1 use F_{v,e}: F_v -> F_e. Please adopt a single convention (vertex-to-edge for the coboundary) and state it explicitly.
[Definition 2] The phrase 'The collections of all ϵ-harmonic edges and harmonic nodes are denoted by...' should read 'ϵ-harmonic edges and ϵ-harmonic nodes' for consistency with Har_0^epsilon(s).
[Framework and Methodology] Minor typo: 'the harmonic set associated with s is defined the union' should be 'defined as the union'.
[Figures] Figures 2 and 4 are illustrative but do not show a concrete harmonic set or an epsilon-harmonic filtration. A small explicit example of Har^epsilon(s) for varying epsilon would help readers understand the construction.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the main theorems are definitional or immediate, but no fitted prediction or self-citation loop is load-bearing.

full rationale

The paper contains no fitted parameters, no empirical predictions, and no load-bearing self-citations. Theorem 1 is an explicit construction: the cellular sheaf is stipulated by setting F_v = F_e = R^d and defining the restriction maps F_{vi⊴{vi,vj}} as scalar multiplication by the attention weight w_ij. This is a modeling relabeling of the GAT triple, not a derivation that uses its own conclusion as a premise. Theorem 2 restates the definition of the global-section space with the observation that all edges are harmonic exactly when C^0_F s = 0, i.e., s ∈ Γ(G;F); it is a direct unpacking of definitions rather than circular inference. Theorem 3 and Corollary 2 define ϵ-harmonic edges by thresholding ∥t_e∥₂, so the monotonicity of the filtration is immediate from the ordering of thresholds. The paper itself concedes in the Future Work section that analyzing the relationship between the proposed sheaf-based harmonic analysis and GDL/TDL learning behavior 'remains to be explored', and it identifies diagnostic use for oversmoothing and heterophily as future work; this is an unsupported-applicability concern, not a circularity. The signal-dependence of actual GAT attention weights means the sheaf is not fixed during training, but Theorem 1 is stated for a frozen triple and does not covertly import that dependence to prove anything. Thus no specific circular reduction can be exhibited.

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

The paper contributes no fitted constants and no new ontological entities. Its formal results rely on standard sheaf theory, the combinatorial Hodge theorem, and the domain assumption that attention weights can be frozen as signal-independent scalar maps. The main conceptual burden is the unsupported step from monotone harmonic filtrations to practical diagnostics for GNN training behavior.

assumptions (4)

standard math Combinatorial Hodge decomposition: ker(L^0_F) = ker(C^0_F) = Γ(G;F).
Invoked in the Mathematical Background section (Eq. 4) to connect the sheaf Laplacian kernel to global sections; this is a standard theorem cited to Hansen and Ghrist (2019).
domain assumption Graphs are finite, undirected, simple, and unweighted.
Stated at the start of Mathematical Background; all definitions, theorems, and the filtration framework are restricted to this setting.
domain assumption A fixed GAT triple (G, (s_vi), W) supplies attention weights as fixed scalar restriction maps.
Theorem 1 treats W as a fixed matrix independent of the node signals; this conflicts with GAT's actual forward pass, where w_ij is a function of node features, so the sheaf is not fixed during training.
standard math The face relation poset identifies a graph with a poset category and a cellular sheaf with a functor to Vect_R.
Used throughout as the definition of cellular sheaves on graphs; standard sheaf theory.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Sheaf-Theoretic and Topological Perspective on Complex Network Modeling and Attention Mechanisms in Graph Neural Models." pith.science (2026). https://pith.science/paper/QHCTWWBM

@misc{pith2026260121207,
  author       = {Pith},
  title        = {Pith review of: A Sheaf-Theoretic and Topological Perspective on Complex Network Modeling and Attention Mechanisms in Graph Neural Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QHCTWWBM}},
  note         = {Machine review of arXiv:2601.21207}
}

read the original abstract

Combinatorial and topological structures, such as graphs, simplicial complexes, and cell complexes, form the foundation of geometric and topological deep learning (GDL and TDL) architectures. These models aggregate signals over such domains, integrate local features, and generate representations for diverse real-world applications. However, the distribution and diffusion behavior of GDL and TDL features during training remains an open and underexplored problem. Motivated by this gap, we introduce a cellular sheaf theoretic framework for modeling and analyzing the local consistency and harmonicity of node features and edge weights in graph-based architectures. By tracking local feature alignments and agreements through sheaf structures, the framework offers a topological perspective on feature diffusion and aggregation. Furthermore, a multiscale extension inspired by topological data analysis (TDA) is proposed to capture hierarchical feature interactions in graph models. This approach enables a joint characterization of GDL and TDL architectures based on their underlying geometric and topological structures and the learned signals defined on them, providing insights for future studies on conventional tasks such as node classification, substructure detection, and community detection.

Figures

Figures reproduced from arXiv: 2601.21207 by the authors.

**Figure 1.** Illustration of a cellular sheaf F : (G, ⊴) → VectR defined on a graph G. Disks on the top layer represent the stalk spaces assigned to the nodes and edges, whose dimensions may vary. In this example, two nodes v and w, together with an edge e, are highlighted along with their corresponding stalk spaces Fv, Fw, and Fe, as well as the restriction maps Fv⊴e and Fw⊴e. graphical illustration of a cellular sheaf define… view at source ↗

**Figure 2.** An illustrative example of a graph G and the constant sheaf R : (G, ⊴) → VectR. The graph G consists of four vertices u, v, w, x and three edges {u, w}, {v, w}, and {w, x}. The second row depicts the graphical correspondences of arbitrary elements s ∈ C 0 (G; R) and t ∈ C 1 (G; R) with the graphical representation of R. The third row depicts the cases where s, s ′ ∈ C 0 (G; R), with s ∈ Γ(G; R) and s ′ ∈/ Γ(G; R). … view at source ↗

**Figure 3.** An example of a graph (G, ⊴) with vertex set V = {u, v, w, x} and edge set E = {{u, w}, {v, w}, {w, x}}. With respect to the partial order ⊴ on G, the Alexandrov open sets Ux, Uw, Uw ∩ Ux, Uw ∪ Ux, and Uu ∪ Ux are shown as the collections of vertices and edges covered by the shaded regions. Framework and Methodology This section elaborates on the theoretical framework underlying the main results (a), (b), and (c), … view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: An illustrative example of the cellular sheaf identification of a GAT triple [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Temporal Sheaf Neural Networks with Dynamic Orthogonal Transport
cs.LG 2026-06 unverdicted novelty 6.0 of 10

TSNN equips temporal graphs with per-node time-varying orthogonal frames, explicit transport, and a geometric-residual decoder, delivering competitive or superior link prediction on benchmarks plus theoretical guarant...

Reference graph

Works this paper leans on

39 extracted references · 5 linked inside Pith · cited by 1 Pith paper

[1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Alexandroff, P. 1937. Diskrete r\" a ume. Matematicheskii Sbornik, 2(3): 501--519

1937
[4]

Arya, S.; Curry, J.; and Mukherjee, S. 2025. A sheaf-theoretic construction of shape space. Foundations of Computational Mathematics, 25(3): 813--863

2025
[5]

S.; and Lio, P

Barbero, F.; Bodnar, C.; de Oc \'a riz Borde, H. S.; and Lio, P. 2022 a . Sheaf Attention Networks. In Workshop on Symmetry and Geometry in Neural Representations at the 36th Conference on Neural Information Processing Systems (NeurIPS 2022)

2022
[6]

Barbero, F.; Bodnar, C.; S\'aez de Oc\'ariz Borde, H.; Bronstein, M.; Veli c kovi\'c, P.; and Li\`o, P. 2022 b . Sheaf Neural Networks with Connection L aplacians. In Topological, Algebraic, and Geometric Learning Workshops 2022, volume 196 of Proceedings of Machine Learning Research (PMLR), 28--36

2022
[7]

Bodnar, C.; Di Giovanni, F.; Chamberlain, B.; Lio, P.; and Bronstein, M. 2022. Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNN s. In Advances in Neural Information Processing Systems (NeurIPS 2022), volume 35, 18527--18541

2022
[8]

F.; Li \'o , P.; and Bronstein, M

Bodnar, C.; Frasca, F.; Wang, Y.; Otter, N.; Montufar, G. F.; Li \'o , P.; and Bronstein, M. 2021. Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks. In Proceedings of the 38th International Conference on Machine Learning (ICML), volume 139, 1026--1037

2021

Show all 39 references

[9]

Bredon, G. E. 1997. Sheaf Theory. United States: Springer New York, NY, 2nd edition. ISBN 978-0-387-94905-5

1997
[10]

M.; Bruna, J.; LeCun, Y.; Szlam, A.; and Vandergheynst, P

Bronstein, M. M.; Bruna, J.; LeCun, Y.; Szlam, A.; and Vandergheynst, P. 2017. Geometric Deep Learning: Going beyond Euclidean data. IEEE Signal Processing Magazine, 34(4): 18--42

2017
[11]

Carlsson, G. 2009. Topology and data. Bulletin of the American Mathematical Society, 46(2): 255--308

2009
[12]

Carlsson, G.; Zomorodian, A.; Collins, A.; and Guibas, L. 2004. Persistence Barcodes for Shapes. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH Symposium on Geometry Processing, 124--135

2004
[13]

Chen, C.; Ye, W.; Zuo, Y.; Zheng, C.; and Ong, S. P. 2019. Graph networks as a universal machine learning framework for molecules and crystals. Chemistry of Materials, 31(9): 3564--3572

2019
[14]

Cohen-Steiner, D.; Edelsbrunner, H.; and Harer, J. 2005. Stability of Persistence Diagrams. In Proceedings of the Twenty-First Annual Symposium on Computational Geometry (SoCG), 263--271

2005
[15]

Cooperband, Z.; and Ghrist, R. 2023. Towards homological methods in graphic statics. Journal of the International Association for Shell and Spatial Structures, 64(4): 266--277

2023
[16]

Cooperband, Z.; and Ghrist, R. 2025. Unified Origami Kinematics via Cosheaf Homology. arXiv:2501.02581

2025 arXiv
[17]

Cooperband, Z.; Ghrist, R.; and Hansen, J. 2023. A Cosheaf Theory of Reciprocal Figures: Planar and Higher Genus Graphic Statics. arXiv:2311.12946

2023 arXiv
[18]

Cooperband, Z.; Lopez, M.; and Schulze, B. 2024. Equivariant Cosheaves and Finite Group Representations in Graphic Statics. arXiv:2401.09392

2024 arXiv
[19]

Curry, J. M. 2014. Sheaves, cosheaves and applications. Ph D thesis, University of Pennsylvania, Philadelphia, United States

2014
[20]

Curry, J. M. 2015. Topological data analysis and cosheaves. Japan Journal of Industrial and Applied Mathematics, 32(2): 333--371

2015
[21]

K.; Maclaurin, D.; Iparraguirre, J.; Bombarell, R.; Hirzel, T.; Aspuru-Guzik, A.; and Adams, R

Duvenaud, D. K.; Maclaurin, D.; Iparraguirre, J.; Bombarell, R.; Hirzel, T.; Aspuru-Guzik, A.; and Adams, R. P. 2015. Convolutional Networks on Graphs for Learning Molecular Fingerprints. In Advances in Neural Information Processing Systems (NIPS 2015), volume 28

2015
[22]

Ebli, S.; Defferrard, M.; and Spreemann, G. 2020. Simplicial Neural Networks. arXiv:2010.03633

2020 arXiv
[23]

Ghrist, R. 2008. Barcodes: T he persistent topology of data. Bulletin of the American Mathematical Society, 45(1): 61--75

2008
[24]

Gong, L.; and Cheng, Q. 2019. Exploiting Edge Features for Graph Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 9203--9211

2019
[25]

Hajij, M.; Istvan, K.; and Zamzmi, G. 2020. Cell Complex Neural Networks. In Topological Data Analysis and Beyond Workshop at the 34th Conference on Neural Information Processing Systems (NeurIPS 2020)

2020
[26]

N.; Birdal, T.; Dey, T

Hajij, M.; Zamzmi, G.; Papamarkou, T.; Miolane, N.; Guzmán-Sáenz, A.; Ramamurthy, K. N.; Birdal, T.; Dey, T. K.; Mukherjee, S.; Samaga, S. N.; Livesay, N.; Walters, R.; Rosen, P.; and Schaub, M. T. 2023. Topological Deep Learning: Going Beyond Graph Data. arXiv:2206.00606

2023 arXiv
[27]

Hansen, J.; and Gebhart, T. 2020. Sheaf Neural Networks. In Topological Data Analysis and Beyond Workshop at the 34th Conference on Neural Information Processing Systems (NeurIPS 2020)

2020
[28]

Hansen, J.; and Ghrist, R. 2019. Toward a spectral theory of cellular sheaves. Journal of Applied and Computational Topology, 3(4): 315--358

2019
[29]

Jiang, B.; Zhang, Z.; Lin, D.; Tang, J.; and Luo, B. 2019. Semi-Supervised Learning With Graph Learning-Convolutional Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

2019
[30]

Kearnes, S.; McCloskey, K.; Berndl, M.; Pande, V.; and Riley, P. 2016. Molecular graph convolutions: moving beyond fingerprints. Journal of Computer-Aided Molecular Design, 30(8): 595---608

2016
[31]

N.; and Welling, M

Kipf, T. N.; and Welling, M. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In Proceedings of the International Conference on Learning Representations (ICLR)

2017
[32]

Li, X.; Sun, L.; Ling, M.; and Peng, Y. 2023. A survey of graph neural network based recommendation in social networks. Neurocomputing, 549: 126441

2023
[33]

Robinson, M. 2014. Topological Signal Processing. Germany: Springer Berlin, Heidelberg. ISBN 978-3-642-36103-6

2014
[34]

Short, R.; Cleveland, J.; Cooperband, Z.; and Moy, M. 2022. On the Current State of Sheaf Theoretic Networking. In 2022 IEEE International Conference on Wireless for Space and Extreme Environments (WiSEE)

2022
[35]

Tennison, B. R. 1975. Sheaf Theory. London Mathematical Society Lecture Note Series. Cambridge University Press. ISBN 978-0-521-20784-3

1975
[36]

Veli c kovi \' c , P.; Cucurull, G.; Casanova, A.; Romero, A.; Li \` o , P.; and Bengio, Y. 2018. Graph Attention Networks. In Proceedings of the International Conference on Learning Representations (ICLR)

2018
[37]

Xie, Q.; Zhu, Y.; Huang, J.; Du, P.; and Nie, J.-Y. 2021. Graph neural collaborative topic model for citation recommendation. ACM Transactions on Information Systems, 40(3): 1--30

2021
[38]

B.; Hayder, Z.; Rolland, V.; Stone, E.; and Petersson, L

Zia, A.; Khamis, A.; Nichols, J.; Tayab, U. B.; Hayder, Z.; Rolland, V.; Stone, E.; and Petersson, L. 2024. Topological deep learning: a review of an emerging paradigm. Artificial Intelligence Review, 57(4): 77

2024
[39]

Zomorodian, A.; and Carlsson, G. 2004. Computing Persistent Homology. In Proceedings of the Twentieth Annual Symposium on Computational Geometry (SoCG), 347--356

2004

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Alexandroff, P. 1937. Diskrete r\" a ume. Matematicheskii Sbornik, 2(3): 501--519

1937

[4] [4]

Arya, S.; Curry, J.; and Mukherjee, S. 2025. A sheaf-theoretic construction of shape space. Foundations of Computational Mathematics, 25(3): 813--863

2025

[5] [5]

S.; and Lio, P

Barbero, F.; Bodnar, C.; de Oc \'a riz Borde, H. S.; and Lio, P. 2022 a . Sheaf Attention Networks. In Workshop on Symmetry and Geometry in Neural Representations at the 36th Conference on Neural Information Processing Systems (NeurIPS 2022)

2022

[6] [6]

Barbero, F.; Bodnar, C.; S\'aez de Oc\'ariz Borde, H.; Bronstein, M.; Veli c kovi\'c, P.; and Li\`o, P. 2022 b . Sheaf Neural Networks with Connection L aplacians. In Topological, Algebraic, and Geometric Learning Workshops 2022, volume 196 of Proceedings of Machine Learning Research (PMLR), 28--36

2022

[7] [7]

Bodnar, C.; Di Giovanni, F.; Chamberlain, B.; Lio, P.; and Bronstein, M. 2022. Neural Sheaf Diffusion: A Topological Perspective on Heterophily and Oversmoothing in GNN s. In Advances in Neural Information Processing Systems (NeurIPS 2022), volume 35, 18527--18541

2022

[8] [8]

F.; Li \'o , P.; and Bronstein, M

Bodnar, C.; Frasca, F.; Wang, Y.; Otter, N.; Montufar, G. F.; Li \'o , P.; and Bronstein, M. 2021. Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks. In Proceedings of the 38th International Conference on Machine Learning (ICML), volume 139, 1026--1037

2021