Pith. sign in

REVIEW 4 major objections 6 minor 63 references

TMetaNet: Topological Meta-Learning Framework for Dynamic Link Prediction

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Per-snapshot learning rates computed from Dowker Zigzag Persistence signatures make dynamic link prediction more accurate and more robust to noise than fixed-weight meta-learning updates.

desk verdict Genuinely new topological representation for dynamic graphs with broad experiments, but the advertised stability theorem does not cover the tested regime and the significance claims outrun the tables. read the letter →

arxiv 2506.00453 v1 pith:LS3XH573 submitted 2025-05-31 cs.LG cs.AI

Hao Li , Hao Wan , Yuzhou Chen , Dongsheng Ye , Yulia Gel , Hao Jiang This is my paper · ORCID
classification cs.LGcs.AI
keywords DowkerZigzagPersistencedynamiclinkpredictionmeta-learninggraphneuralnetworkspersistenthomologyε-netlandmarkselectionnoiserobustness
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

TMetaNet sets out to show that the evolving high-order shape of a graph — which cycles and clusters appear and disappear as the network changes — is the right signal for deciding how fast a dynamic graph neural network should adapt at each moment. Instead of the fixed weight-update parameters used by ROLAND and the fixed learning rate used by WinGNN, TMetaNet computes a per-snapshot learning rate from topological signatures called Dowker Zigzag Persistence (DZP), which track high-order features across snapshot windows while using only a small landmark subset of nodes to keep the computation affordable. The paper argues that this topological guidance makes dynamic link prediction more accurate (average gains of 3.29% in ACC and 5.52% in MRR under the ROLAND live-update setting, and 3.27% and 24.08% under the WinGNN future-forecast setting, with gains up to 74.70% on individual benchmarks) and considerably more resilient to edge-flipping noise (up to 31.1% under poisoning attacks). If correct, this establishes a broader principle: computable topological summaries can serve as task descriptors for meta-learning on structured, time-varying data.

What carries the argument

The load-bearing object is the Dowker Zigzag Persistence signature: a way of running zigzag persistence on a dynamic graph using only a small subset of its nodes. For each snapshot window, a Dowker complex is built from a landmark set selected by an $\epsilon$-net algorithm — seeded with the previous snapshot's landmarks so the sets stay consistent over time — while the remaining nodes serve as witnesses that decide which simplices appear at scale $\delta$; the complexes of consecutive snapshots and their unions are chained into a zigzag sequence, reduced to persistence diagrams, and vectorized into zigzag persistence images (ZPI). The mechanism that carries the argument is the Topological Learning Rate Adaptor: the difference between the ZPIs of adjacent snapshots passes through a residual CNN and fully connected layers to produce the scalar learning rate $r$, which replaces the fixed update weight in $w_{t+1} = w_t - \eta \cdot r \cdot \nabla_w \mathcal{L}(w_t)$. This is what lets high-order structural change, rather than elapsed time, set the step size of the model's adaptation, and the landmark restriction is what keeps the signature affordable (the claimed 0-dimensional complexity falls from $O(m \log^2 N + m \log m)$ to $O(m \log^2 n' + m \log m)$, where $n'$ is the number of landmarks).

What would settle it

Construct two $\epsilon$-interleaved dynamic graphs whose $\epsilon$-net landmark selections fall in different connected regions, so the landmark complexes are no longer interleaved; if their DZP bottleneck distance exceeds $K \cdot d_G$, the stability bound fails and the theoretical basis of the noise-robustness claim collapses. As an empirical companion, re-run the two headline results with the released code — the 74.70% MRR gain on Reddit-Body and the 31.1% robustness gain on Reddit-Title — against WinGNN and ROLAND on identical splits with all baselines retuned.

Watch

Extended reading notes

Core claim

The paper's central claim is that high-order topological features of a dynamic graph, tracked through time as Dowker Zigzag Persistence signatures, carry information about how the graph is evolving that timestamps and fixed update rules throw away, and that feeding this information into the optimizer improves prediction. Concretely, TMetaNet builds, for each snapshot window, a zigzag persistence diagram over Dowker complexes constructed from a small $\epsilon$-net landmark set (with the remaining nodes acting as witnesses), vectorizes the diagrams into persistence images, differences the images of adjacent snapshots, and maps that difference through a CNN and fully connected layers to a scalar learning rate $r$, applied as $w_{t+1} = w_t - \eta \cdot r \cdot \nabla_w \mathcal{L}(w_t)$. The paper also claims a stability theorem: for two structurally close ($\epsilon$-interleaved) dynamic graphs the bottleneck distance between their DZP diagrams is bounded by a constant times their structural distance, $d_B(\mathrm{Dgm}^{\mathrm{DZP}}(G_X), \mathrm{Dgm}^{\mathrm{DZP}}(G_Y)) \le K \cdot d_G(G_X, G_Y)$, which it offers as the reason the topological learning rates remain reliable when the graph is perturbed. Empirically, the claim is that TMetaNet outperforms the fixed-weight baselines on six real networks under both evaluation protocols and degrades more gracefully than WinGNN under evasion and poisoning attacks.

Load-bearing premise

The load-bearing premise is that two dynamic graphs that are structurally close, in the paper's $\epsilon$-interleaving sense, necessarily have close Dowker Zigzag Persistence diagrams; the proof assumes that $\epsilon$-net landmark sets inherit the interleaving of the full graphs and that zigzag persistence obeys the same bottleneck-stability theorem as ordinary single-parameter persistence, a step the appendix supports only with a sketch ('the same principle holds').

Editorial extensions

If this is right

  • Dynamic link prediction improves over the fixed-weight meta-learning baselines: on average, TMetaNet gains 3.29% in ACC and 5.52% in MRR under the ROLAND live-update setting, and 3.27% in ACC and 24.08% in MRR under the WinGNN future-forecast setting, with statistically significant differences ($p < 0.05$) on most datasets.
  • The benefit comes from the topological content of the signature rather than from having an extra input vector: replacing the DZP images with random images, or with images matched only in statistical distribution, lowers accuracy and MRR in the ablations, and the random replacement hurts the most.
  • The gains concentrate where graphs genuinely change and where the task is future prediction: on more homogeneous datasets such as Reddit-Body under the live-update setting the improvement is small or negative, which the paper reads as the topological signal being strongest under pronounced structural dynamics.
  • Landmark subsampling makes the zigzag machinery practical: the DZP complexity bound replaces the full node count $N$ with the $\epsilon$-net size $n'$ (0-dimensional complexity falls from $O(m \log^2 N + m \log m)$ to $O(m \log^2 n' + m \log m)$), and on Bitcoin-Alpha the Dowker construction cuts per-snapshot complex-building overhead by about 46% relative to Vietoris-Rips-based zigzag persistence
  • If Theorem 4.5 holds, small structural perturbations shift the DZP diagrams only a little, which the paper connects to the observed robustness: on Reddit-Title, TMetaNet degrades more gracefully than WinGNN under both evasion and poisoning edge-flip attacks, with up to a 31.1% MRR advantage in the WinGNN setting.

Reading between the lines

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

  • A natural extension is to other dynamic-graph tasks — node classification, anomaly detection, streaming community detection — wherever the right adaptation speed depends on structural change rather than on time alone, since the DZP signature is computed once and can feed any model's update rule.
  • The landmark-selection rule is a testable lever the paper leaves unturned: DZP seeds landmarks by node degree and by overlap with the previous snapshot, so varying the selection rule (random, betweenness-driven, or uniform) would reveal whether the predictive signal lives specifically in central nodes or in the $\epsilon$-net property itself.
  • The stability bound, if it holds as stated, suggests a certification route the paper does not pursue: because $d_B \le K \cdot d_G$ limits how much noise can move the topological signature, one could convert the empirical robustness into a per-instance guarantee on how far an adversarial edge edit can shift the learned learning rate.
  • The distribution-matched ablation still beating random replacement hints that part of the gain is statistical rather than strictly topological; replacing DZP with other structured but non-topological inputs, such as degree sequences or spectral summaries, would sharpen the claim that topology specifically is what matters.
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 / 6 minor

Summary. The paper proposes Dowker Zigzag Persistence (DZP), a topological signature for dynamic graphs constructed from Dowker complexes on small landmark sets and zigzag persistence, and combines it with a meta-learning framework, TMetaNet, that learns a per-snapshot learning rate from differences between zigzag persistence images. The central theoretical claim is Theorem 4.5, which asserts that the bottleneck distance between DZP diagrams of two dynamic graphs is bounded by a constant times a structural graph discrepancy metric dG. The paper further reports experiments on six dynamic link prediction datasets under two training settings, claiming state-of-the-art accuracy/MRR and noise robustness, with ablations, sensitivity analyses, and a code repository.

Significance. If the main empirical claim holds, the idea of using high-order topological signatures of graph snapshots to adapt meta-learning parameter updates is a plausible and potentially useful contribution to dynamic graph learning. The paper has concrete strengths: a public codebase, experiments across six datasets, multiple evaluation metrics, ablation studies, noise robustness tests, and a comparison of running times. However, the significance is substantially moderated by the state of the theoretical contribution: the stability theorem is advertised as a main result, yet the supplied proof has load-bearing gaps and is not applicable to the parameter regime used in the experiments. The empirical claims also need more careful statistical support, since no p-values are reported and at least one headline comparison is not a win. The contribution is therefore best assessed as promising but requiring a thorough revision of the theory and the statistical reporting.

major comments (4)
  1. [Theorem 4.5, Appendix B, Theorem B.9, Appendix D.3] The stability theorem is not proved for the parameter regime used in the paper. Theorem B.9 requires ε ≥ ϵ + δmax, where δmax is the maximum diameter of the smoothed vertex sets, but Appendix D.3 states that all experiments use ε=1 and δ=1. The graphs in Table 3 are large real-world networks whose smoothed snapshots have diameters far larger than 1, so the hypothesis of Theorem B.9 is violated in the evaluated regime. Conversely, if ε were increased to satisfy the condition, the ε-net would contain at most one landmark per connected component, making the Dowker complex too sparse to carry the higher-order features that DZP is designed to capture. Thus Theorem 4.5 cannot support the DZP robustness claims in the setting where TMetaNet is tested.
  2. [Appendix B, Theorem B.10, proof step 1] The proof asserts a δ-interleaving of the Dowker complexes D(L_t^X) and D(L_t^Y) but never constructs a simplicial map between them. The proposed map Φ_t is described as a set of simplices in D(L_t^Y) with a witness condition, and it is not shown to be a well-defined simplicial map. No correspondence between the witness sets W_t^X and W_t^Y is established, and the structural metric dG in Definition B.8 has terms for vertex coverage and edge-weight discrepancies but no term that controls distances between witnesses or how shortest-path distances change under the tripod. Consequently, the claimed bound dB ≤ K·dG does not follow from the supplied argument, and the robustness results in §6.4 cannot be attributed to the proved stability of DZP.
  3. [Appendix B, Theorem B.10, proof step 3] The final step from an interleaving of complexes to a bottleneck bound for zigzag persistence is asserted rather than derived. The text says that for zigzag modules 'the same principle holds' and then absorbs all technicalities into an unspecified dimension-dependent constant K′. For ordinary one-parameter persistence the implication is the classical stability theorem, but for zigzag persistence one must explicitly construct an interleaving of the zigzag modules and either invoke a known stability theorem for zigzag modules or prove the needed statement. Neither is supplied nor cited, so Eq. (23) is not a derivation.
  4. [Section 6, first paragraph; Table 1; Table 5] The claim that TMetaNet achieves 'statistically significant improvements (p < 0.05)' is not supported by the paper: no p-values, test statistics, or multiple-comparison procedures are reported anywhere. Moreover, Table 1 shows TMetaNet under the ROLAND setting is worse than ROLAND on Reddit-Body ACC (89.59 ± 1.17 vs 91.63 ± 0.09), and Table 5 shows AUC losses on Reddit-Body and UCI under the same setting. The blanket significance claim should be restricted to the specific comparisons where pairwise tests were actually performed, and the test details should be reported.
minor comments (6)
  1. [Definition 4.1] There is an inconsistency in the notation: the Dowker complex is defined with a distance threshold δ in the set-builder expression, while the immediately following text says 'ε represents the maximum allowable distance'. Please use a single symbol consistently.
  2. [Definition 4.4] The displayed formula for dG has a set-braces expression whose contents are not fully visible in the main text (the terms after 'max { ' are missing). The appendix version is clearer; please fix the main-text rendering.
  3. [Appendix C.1, Algorithm 2] In Algorithm 2, for t > 1 the algorithm first sets landmarks_t from the highest-degree nodes of the current snapshot and only later uses the mapped landmark seed for witnesses; this is inconsistent with the text in Section 5.1, which says ε-seed ensures that landmark sets of adjacent snapshots remain consistent. Please clarify how the seed-based expansion is actually applied.
  4. [Section 6, abstract] The headline '74.70%' and '31.1%' improvements are relative improvements over a specific baseline in a specific setting (Table 1, Reddit-Body MRR under the WinGNN setting, and Figure 4). The abstract and contribution list should state that these are relative gains with the exact baseline and setting, to avoid implying absolute improvements of this magnitude.
  5. [Appendix D.3] The grid search ranges for the meta-learning rate and dropout rate are not specified, and no seed or software-version details are given. Reporting those ranges would improve reproducibility.
  6. [Throughout] There are several typographical errors, including 'pre-trainng' in the Related Work, 'Leveraging the mathematical the theory' in Section 3, and 'discusson' in Section 4. These should be corrected in a revision.

Circularity Check

2 steps flagged · score 4.0 of 10

Moderate partial circularity: the advertised DZP stability theorem (Theorem 4.5) is inherited from the same authors' prior Theorem 3 in Ye et al. (2023) with the DZP-specific interleaving asserted, while the benchmark link-prediction results are measured and independent.

  1. self citation load bearing [Theorem 4.5 (Section 4) and Appendix B, opening paragraph and Theorem B.10 proof]
    "The proof is based on the Theorem 3 in (Ye et al., 2023), showing that DZP remains stable under controlled perturbations of the underlying graph structure."

    Theorem 4.5 is advertised as a new theoretical contribution (Contribution 2: 'We derive theoretical stability guarantees of the resulting DZP representations'), but the paper's own appendix attributes the target result to Theorem 3 of Ye et al. (2023), whose authors (Ye, Jiang, Li) overlap three co-authors of the present paper. The DZP-specific bridge is asserted rather than established: Theorem B.9's ε-net inheritance carries a stated hypothesis (ε ≥ ϵ + diam(SϵVt)) that is not shown to hold in the paper's own evaluation setting ε=1, δ=1 (Appendix D.3), and the transfer from landmark interleavings to Dowker complexes with witnesses is declared without derivation. The structural metric dG itself is imported from the same prior work ('inspired by Ye et al. (2023)').

  2. other [Appendix B, Theorem B.10 proof (steps 1–3), Definition 4.1, and Theorem B.9]
    "Φt(σ) = {τ ∈ D(LY t) | ∃witness w : d(σ, w) ≤ δ} This guarantees: D(LX t) ,→ D(LY t)δ and D(LY t) ,→ D(LX t)δ ... While conceptually more involved, the same principle holds: if we have a δ-interleaving between two zigzag modules (compatible inclusions that shift by at most δ), then the birth and death of any homology class can shift by at most δ. Hence the bottleneck distance of the two zigzag persistence modules is at most δ."

    The proof defines the asserted interleaving map Φt as 'has a witness within δ', which is exactly Definition 4.1's Dowker complex membership condition ('∃w ∈ Wt such that ∀l ∈ σ, d(l, w) ≤ δ'). The δ-interleaving between the two Dowker zigzag modules is then declared 'guaranteed' at the same scale δ used to construct the complexes, so the bottleneck scale is inherited from the construction's own parameter rather than from the graph discrepancy dG. The final step from interleaving to bottleneck distance is assumed via 'the same principle holds' with only the single-parameter Cohen-Steiner et al. (2005) result cited and no zigzag stability theorem referenced, and the dimension-dependent constant K' is 'absorbed' without derivation.

full rationale

The paper separates into an empirical chain and a theoretical chain. The empirical chain — Table 1 results against public benchmarks (GCN, EvolveGCN, ROLAND, WinGNN, DeGNN) under two published training settings, the Section 6.2 ablations replacing ZPI content with Random and Dist controls, and the Section 6.4 noise-injection measurements — is directly measured on held-out test slices or future-period splits; none of these numbers is produced from Theorem 4.5 or from a fitted parameter renamed as a prediction, so the headline SOTA and robustness claims would stand even if the stability theorem were false. The circularity burden is confined to the theoretical chain. Contribution 2 announces 'We derive theoretical stability guarantees', but Appendix B declares the proof 'is based on the Theorem 3 in (Ye et al., 2023), showing that DZP remains stable'; Ye et al. (2023) shares three co-authors (Ye, Jiang, Li) with this paper, and the DZP-specific bridge is the asserted δ-interleaving whose witness condition is Definition 4.1's membership condition, followed by the asserted zigzag bottleneck principle ('the same principle holds') with a dimension constant K' that is 'absorbed' rather than computed; Theorem B.9's hypothesis ε ≥ ϵ + diam(SϵVt) is not shown to hold for the ε=1, δ=1 settings of Appendix D.3. The appendix itself signals these asserted steps, which I weigh as in-scope evidence per the review rules. The theory claim is therefore load-bearing self-citation with asserted content, justifying a mid-scale score; because no empirical prediction reduces by construction, the score is 4 rather than 6 or higher.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper does not introduce new physical or mathematical entities; DZP is a computational construction, and landmark/witness sets come from the Dowker complex literature. The main unstated costs are the tuned hyperparameters and the assumed stability of zigzag persistence under interleaving.

free parameters (6)
  • epsilon (ε-net radius) = 1
    Chosen for all final runs (Appendix D.3); sensitivity analysis in Section 6.3 tested (ε,δ) values from 1 to 4 and found dataset-dependent variation.
  • delta (Dowker complex radius) = 1
    Set together with ε in Appendix D.3; it controls which witness nodes create simplices in the Dowker complex.
  • window size w = full
    Final setting per Appendix D.3; Section 6.3 shows MRR varies strongly with w in the ROLAND setting, so this is a free choice rather than a fixed constant.
  • ZPI image size = 50
    Persistence image grid resolution chosen in Appendix D.3; not swept.
  • meta-lr = not reported
    Selected by per-dataset grid search (Appendix D.3), values not listed; it directly controls the learning-rate adaptor.
  • dropout rate = not reported
    Selected by grid search per dataset; affects generalization of the adaptor.
assumptions (4)
  • domain assumption Dowker complex on ε-net landmarks preserves enough topology for zigzag persistence.
    Used throughout Sections 4 and 5; the complexity bound and stability theorem both depend on this, but no formal guarantee is proven for the specific ε-seed algorithm.
  • domain assumption The structural metric dG and ε-interleaving framework from Ye et al. (2023) is valid for DZP.
    Definition 4.4 and Theorem 4.5 adopt dG and interleaving from a same-group prior paper; the present paper does not reprove these foundations.
  • ad hoc to paper Zigzag persistence modules satisfy bottleneck stability under δ-interleaving.
    Appendix B assumes the classical stability argument extends to zigzag modules ('the same principle holds'); this is the key unproven step in Theorem 4.5.
  • domain assumption ε-net cardinality bound n' = n/(ε+1)+2n/d holds for the ε-seed algorithm.
    Section 4 cites Choi et al. (2024) for this bound; the algorithm in Appendix C.1 may not satisfy the exact conditions of that bound.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TMetaNet: Topological Meta-Learning Framework for Dynamic Link Prediction." pith.science (2026). https://pith.science/paper/LS3XH573

@misc{pith2026250600453,
  author       = {Pith},
  title        = {Pith review of: TMetaNet: Topological Meta-Learning Framework for Dynamic Link Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LS3XH573}},
  note         = {Machine review of arXiv:2506.00453}
}
read the original abstract

Dynamic graphs evolve continuously, presenting challenges for traditional graph learning due to their changing structures and temporal dependencies. Recent advancements have shown potential in addressing these challenges by developing suitable meta-learning-based dynamic graph neural network models. However, most meta-learning approaches for dynamic graphs rely on fixed weight update parameters, neglecting the essential intrinsic complex high-order topological information of dynamically evolving graphs. We have designed Dowker Zigzag Persistence (DZP), an efficient and stable dynamic graph persistent homology representation method based on Dowker complex and zigzag persistence, to capture the high-order features of dynamic graphs. Armed with the DZP ideas, we propose TMetaNet, a new meta-learning parameter update model based on dynamic topological features. By utilizing the distances between high-order topological features, TMetaNet enables more effective adaptation across snapshots. Experiments on real-world datasets demonstrate TMetaNet's state-of-the-art performance and resilience to graph noise, illustrating its high potential for meta-learning and dynamic graph analysis. Our code is available at https://github.com/Lihaogx/TMetaNet.

Figures

Figures reproduced from arXiv: 2506.00453 by the authors.

Figure 1
Figure 1. The parameter update mechanisms based on meta-learning in ROLAND, WinGNN, and TMetaNet. (a) In ROLAND, model parameters between adjacent time slices are updated using fixed meta-learning weights. (b) WinGNN performs parameter updates between adjacent time slices with a fixed learning rate, incorporating a window gradient aggregation mechanism to replace explicit time encoding, making the embedding Et invisible to th… view at source ↗
Figure 2
Figure 2. The overview of TMetaNet. 5.2. Topological Learning Rate Adaptor For adjacent snapshots, e.g., Gt and Gt+1, the Topologi￾cal Signature Generator computes their corresponding ZPI sequences Zn,k(Gt) and Zn,k(Gt+1). Subsequently, the Topological Learning Rate Adaptor extracts the difference between adjacent ZPIs through a convolutional neural net￾work, and uses this as input to learn a scalar through fully connected la… view at source ↗
Figure 3
Figure 3. Impact of hyperparameters of zigzag persistence. The top figure shows the results under the ROLAND setting, while the bottom figure corresponds to the WinGNN setting. that longer temporal aggregation dilutes informative signals and introduces noise. In the WinGNN setting, the perfor￾mance is generally more stable and less sensitive to window size. The (1, 1) configuration again achieves the highest peak MRR at w = 1… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Robustness analysis of TMetaNet under different settings, where e 10 represents a 10% edge evasion attack. The top figure shows the results under the ROLAND setting, while the bottom figure corresponds to the WinGNN setting. 7. Conclusion We have presented Dowker Zigza…
Figure 5
Figure 5. Figure 5: illustrates a comparative analysis of distinct task splitting methodologies. ROLAND implements a snapshot-wise partitioning strategy, wherein each temporal snapshot Gt is partitioned into training, validation, and testing subsets. The model undergoes training on the tr…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 61 canonical work pages

  1. [1]

    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 format.date year duplicate empty "emp...

  2. [2]

    G., Bennink, R., Chen, Y., Fr \' as, J., Gel, Y

    Aksoy, S. G., Bennink, R., Chen, Y., Fr \' as, J., Gel, Y. R., Kay, B., Naumann, U., Marrero, C. O., Petyuk, A. V., Roy, S., et al. Seven open problems in applied combinatorics. Journal of Combinatorics, 14 0 (4): 0 559--601, 2023

    work page 2023

  3. [3]

    A., Basu, D., and Bressan, S

    Arafat, N. A., Basu, D., and Bressan, S. Topological data analysis with -net induced lazy witness complex. In International Conference on Database and Expert Systems Applications, pp.\ 376--392, 2019

    work page 2019

  4. [4]

    A., Basu, D., Gel, Y., and Chen, Y

    Arafat, N. A., Basu, D., Gel, Y., and Chen, Y. When witnesses defend: A witness graph topological layer for adversarial graph learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp.\ 15408--15416, 2025

    work page 2025

  5. [5]

    Meta-Graph: Few Shot Link Prediction via Meta Learning

    Bose, A. J., Jain, A., Molino, P., and Hamilton, W. L. Meta-Graph : Few shot link prediction via meta learning. arXiv Preprint arXiv:1912.09867, 2019

    work page Pith review arXiv 1912

  6. [6]

    and Gabrielsson, R

    Carlsson, G. and Gabrielsson, R. B. Topological approaches to deep learning. In Topological Data Analysis: The Abel Symposium 2018, pp.\ 119--146, 2020

    work page 2018

  7. [7]

    and Silva, V

    Carlsson, G. and Silva, V. Zigzag persistence. Foundations of Computational Mathematics, 10: 0 367–405, 2010

    work page 2010

  8. [8]

    Parametrized homology via zigzag persistence

    Carlsson, G., De Silva, V., Kali s nik, S., and Morozov, D. Parametrized homology via zigzag persistence. Algebraic & Geometric Topology, 19 0 (2): 0 657--700, 2019

    work page 2019

Show all 63 references
  1. [9]

    Few-shot learning on graphs via super-classes based on graph spectral measures

    Chauhan, J., Nathani, D., and Kaul, M. Few-shot learning on graphs via super-classes based on graph spectral measures. In International Conference on Learning Representations, 2020

    2020

  2. [10]

    and Michel, B

    Chazal, F. and Michel, B. An introduction to topological data analysis: fundamental and practical aspects for data scientists. Frontiers in Artificial Intelligence, 4: 0 667963, 2021

    2021

  3. [11]

    Persistence stability for geometric complexes

    Chazal, F., De Silva, V., and Oudot, S. Persistence stability for geometric complexes. Geometriae Dedicata, 173 0 (1): 0 193--214, 2014

    2014

  4. [12]

    and Zhang, A

    Chen, J. and Zhang, A. HetMAML : Task-heterogeneous model-agnostic meta-learning for few-shot learning across modalities. In Proceedings of the 30th ACM International Conference on Information and Knowledge Management, pp.\ 191--200, 2021

    2021

  5. [13]

    and Gel, Y

    Chen, Y. and Gel, Y. Topological zigzag spaghetti for diffusion-based generation and prediction on graphs. In International Conference on Learning Representations, 2025

    2025

  6. [14]

    Topological relational learning on graphs

    Chen, Y., Coskunuzer, B., and Gel, Y. Topological relational learning on graphs. In Advances in Neural Information Processing Systems, volume 34, pp.\ 27029--27042, 2021 a

    2021

  7. [15]

    Chen, Y., Segovia, I., and Gel, Y. R. Z-GCNETs : Time zigzags at graph convolutional networks for time series forecasting. In International Conference on Machine Learning, pp.\ 1684--1694, 2021 b

    2021

  8. [16]

    Chen, Y., Gel, Y., and Poor, H. V. Time-conditioned dances with simplicial complexes: Zigzag filtration curve based supra-hodge convolution networks for time-series forecasting. In Advances in Neural Information Processing Systems, volume 35, pp.\ 8940--8953, 2022 a

    2022

  9. [17]

    TAMP-S2GCNets : coupling time-aware multipersistence knowledge representation with spatio-supra graph convolutional networks for time-series forecasting

    Chen, Y., Segovia-Dominguez, I., Coskunuzer, B., and Gel, Y. TAMP-S2GCNets : coupling time-aware multipersistence knowledge representation with spatio-supra graph convolutional networks for time-series forecasting. In International Conference on Learning Representations, 2022 b

    2022

  10. [18]

    Chen, Y., Sizikova, E., and Gel, Y. R. Topoattn-nets: Topological attention in graph representation learning. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pp.\ 309--325, 2022 c

    2022

  11. [19]

    Chen, Y., Heleno, M., Moreira, A., and Gel, Y. R. Topological graph convolutional networks solutions for power distribution grid planning. In Pacific-Asia Conference on Knowledge Discovery and Data Mining, pp.\ 123--134, 2023

    2023

  12. [20]

    W., Chen, Y., Fr \' as, J., Castillo, J., and Gel, Y

    Choi, J. W., Chen, Y., Fr \' as, J., Castillo, J., and Gel, Y. Revisiting link prediction with the dowker complex. In Pacific-Asia Conference on Knowledge Discovery and Data Mining, pp.\ 418--430, 2024

    2024

  13. [21]

    and M \'e moli, F

    Chowdhury, S. and M \'e moli, F. A functorial dowker theorem and persistent homology of asymmetric networks. Journal of Applied and Computational Topology, 2: 0 115--175, 2018

    2018

  14. [22]

    The importance of forgetting: Limiting memory improves recovery of topological characteristics from neural data

    Chowdhury, S., Dai, B., and M \'e moli, F. The importance of forgetting: Limiting memory improves recovery of topological characteristics from neural data. PloS One, 13 0 (9): 0 e0202561, 2018

    2018

  15. [23]

    Stability of persistence diagrams

    Cohen-Steiner, D., Edelsbrunner, H., and Harer, J. Stability of persistence diagrams. In Proceedings of the Twenty-First Annual Symposium on Computational Geometry, pp.\ 263--271, 2005

    2005

  16. [24]

    Coskunuzer, B., Segovia-Dominguez, I., Chen, Y., and Gel, Y. R. Time-aware knowledge representations of dynamic objects with multidimensional persistence. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 11678--11686, 2024

    2024

  17. [25]

    and Carlsson, G

    De Silva, V. and Carlsson, G. E. Topological estimation using witness complexes. In Symposium on Point Based Graphics, pp.\ 157--166, 2004

    2004

  18. [26]

    Dey, T. K. and Hou, T. Computing zigzag persistence on graphs in near-linear time. In 37th International Symposium on Computational Geometry, pp.\ 30--1, 2021

    2021

  19. [27]

    Dey, T. K. and Hou, T. Computing zigzag vineyard efficiently including expansions and contractions. In 40th International Symposium on Computational Geometry, pp.\ 49--1, 2024

    2024

  20. [28]

    A comprehensive survey of dynamic graph neural networks: Models, frameworks, benchmarks, experiments and challenges

    Feng, Z., Wang, R., Wang, T., Song, M., Wu, S., and He, S. A comprehensive survey of dynamic graph neural networks: Models, frameworks, benchmarks, experiments and challenges. arXiv Preprint arXiv:2405.00476, 2024

    2024 arXiv

  21. [29]

    Model-Agnostic Meta-Learning for fast adaptation of deep networks

    Finn, C., Abbeel, P., and Levine, S. Model-Agnostic Meta-Learning for fast adaptation of deep networks. In International Conference on Machine Learning, pp.\ 1126--1135, 2017

    2017

  22. [30]

    Gharoun, H., Momenifar, F., Chen, F., and Gandomi, A. H. Meta-learning approaches for few-shot learning: A survey of recent advances. ACM Computing Surveys, 56 0 (12): 0 1--41, 2024

    2024

  23. [31]

    Ghrist, R. W. Elementary applied topology, volume 1. Createspace, 2014

    2014

  24. [32]

    A survey of topological machine learning methods

    Hensel, F., Moor, M., and Rieck, B. A survey of topological machine learning methods. Frontiers in Artificial Intelligence, 4: 0 681108, 2021

    2021

  25. [33]

    Going beyond persistent homology using persistent homology

    Immonen, J., Souza, A., and Garg, V. Going beyond persistent homology using persistent homology. In Advances in Neural Information Processing Systems, volume 36, pp.\ 63150--63173, 2023

    2023

  26. [34]

    and Russold, F

    Kerber, M. and Russold, F. Graphcode: Learning from multiparameter persistent homology using graph neural networks. In Advances in Neural Information Processing Systems, volume 37, pp.\ 41103--41131, 2024

    2024

  27. [35]

    Analysis of dynamic graphs and dynamic metric spaces via zigzag persistence

    Kim, W., Memoli, F., and Smith, Z. Analysis of dynamic graphs and dynamic metric spaces via zigzag persistence. In Topological Data Analysis: The Abel Symposium 2018, pp.\ 371--389, 2020

    2018

  28. [36]

    Edge weight prediction in weighted signed networks

    Kumar, S., Spezzano, F., Subrahmanian, V., and Faloutsos, C. Edge weight prediction in weighted signed networks. In 2016 IEEE 16th International Conference on Data Mining, pp.\ 221--230, 2016

    2016

  29. [37]

    L., Leskovec, J., and Jurafsky, D

    Kumar, S., Hamilton, W. L., Leskovec, J., and Jurafsky, D. Community interaction and conflict on the web. In Proceedings of the 2018 World Wide Web Conference, pp.\ 933--943, 2018 a

    2018

  30. [38]

    REV2 : Fraudulent user prediction in rating platforms

    Kumar, S., Hooi, B., Makhija, D., Kumar, M., Faloutsos, C., and Subrahmanian, V. REV2 : Fraudulent user prediction in rating platforms. In Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, pp.\ 333--341, 2018 b

    2018

  31. [39]

    Dynamic neural dowker network: Approximating persistent homology in dynamic directed graphs

    Li, H., Jiang, H., Jiajun, F., Ye, D., and Du, L. Dynamic neural dowker network: Approximating persistent homology in dynamic directed graphs. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp.\ 1554--1564, 2024

    2024

  32. [40]

    R., and Kantarcioglu, M

    Li, Y., Islambekov, U., Akcora, C., Smirnova, E., Gel, Y. R., and Kantarcioglu, M. Dissecting ethereum blockchain analytics: What we learn from topology and geometry of the ethereum graph? In Proceedings of the 2020 SIAM International Conference on Data Mining, pp.\ 523--531, 2020

    2020

  33. [41]

    R., and Chen, Y

    Liang, P., Gel, Y. R., and Chen, Y. Topology-informed pre-training of graph neural networks. In IEEE International Conference on Acoustics, Speech and Signal Processing, pp.\ 1--5, 2025

    2025

  34. [42]

    Dowker complex based machine learning (dcml) models for protein-ligand binding affinity prediction

    Liu, X., Feng, H., Wu, J., and Xia, K. Dowker complex based machine learning (dcml) models for protein-ligand binding affinity prediction. PLoS Computational Biology, 18 0 (4): 0 e1009943, 2022

    2022

  35. [43]

    M., and Khalil, A

    Ma, T., Su, Y., Abdel Wahab, M. M., and Khalil, A. A. E. ZPDSN : spatio-temporal meteorological forecasting with topological data analysis. Applied Intelligence, 55 0 (1): 0 9, 2025

    2025

  36. [44]

    Dynamic graph convolutional networks

    Manessi, F., Rozza, A., and Manzo, M. Dynamic graph convolutional networks. Pattern Recognition, 97: 0 107000, 2020

    2020

  37. [45]

    Zigzag persistent homology for processing neuronal images

    Mata, G., Morales, M., Romero, A., and Rubio, J. Zigzag persistent homology for processing neuronal images. Pattern Recognition Letters, 62: 0 55--60, 2015

    2015

  38. [46]

    J., and Purvine, E

    Myers, A., Bittner, A., Aksoy, S., Best, D., Henselman-Petrusek, G., Jenne, H., Joslyn, C., Kay, B., Seppala, G., Young, S. J., and Purvine, E. Malicious cyber activity detection using zigzag persistence. In 2023 IEEE Conference on Dependable and Secure Computing, pp.\ 1--8, 2023 a

    2023

  39. [47]

    A., and Munch, E

    Myers, A., Mu \ n oz, D., Khasawneh, F. A., and Munch, E. Temporal network analysis using zigzag persistence. EPJ Data Science, 12 0 (1): 0 6, 2023 b

    2023

  40. [48]

    Panzarasa, P., Opsahl, T., and Carley, K. M. Patterns and dynamics of users' behavior and interaction: Network analysis of an online community. Journal of the American Society for Information Science and Technology, 60 0 (5): 0 911--932, 2009

    2009

  41. [49]

    EvolveGCN : Evolving graph convolutional networks for dynamic graphs

    Pareja, A., Domeniconi, G., Chen, J., Ma, T., Suzumura, T., Kanezashi, H., Kaler, T., Schardl, T., and Leiserson, C. EvolveGCN : Evolving graph convolutional networks for dynamic graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pp.\ 5363--5370, 2020

    2020

  42. [50]

    and Pan, S

    Peng, D. and Pan, S. J. Clustered task-aware meta-learning by learning from learning paths. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45 0 (8): 0 9426--9438, 2023

    2023

  43. [51]

    S., Lee, S

    Pun, C. S., Lee, S. X., and Xia, K. Persistent-homology-based machine learning: a survey and a comparative study. Artificial Intelligence Review, 55 0 (7): 0 5169--5213, 2022

    2022

  44. [52]

    and Carlsson, G

    Tausz, A. and Carlsson, G. Applications of zigzag persistence to topological data analysis. arXiv Preprint arXiv:1108.3545, 2011

    2011 arXiv

  45. [53]

    Advances and challenges in meta-learning: A technical review

    Vettoruzzo, A., Bouguelia, M.-R., Vanschoren, J., R \"o gnvaldsson, T., and Santosh, K. Advances and challenges in meta-learning: A technical review. IEEE transactions on pattern analysis and machine intelligence, 46 0 (7): 0 4763--4779, 2024

    2024

  46. [54]

    Link prediction with persistent homology: An interactive view

    Yan, Z., Ma, T., Gao, L., Tang, Z., and Chen, C. Link prediction with persistent homology: An interactive view. In International Conference on Machine Learning, pp.\ 11659--11669, 2021

    2021

  47. [55]

    Few-shot link prediction in dynamic networks

    Yang, C., Wang, C., Lu, Y., Gong, X., Shi, C., Wang, W., and Zhang, X. Few-shot link prediction in dynamic networks. In Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining, pp.\ 1245--1255, 2022

    2022

  48. [56]

    Graph few-shot learning via knowledge transfer

    Yao, H., Zhang, C., Wei, Y., Jiang, M., Wang, S., Huang, J., Chawla, N., and Li, Z. Graph few-shot learning via knowledge transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pp.\ 6656--6663, 2020

    2020

  49. [57]

    Stable distance of persistent homology for dynamic graph comparison

    Ye, D., Jiang, H., Jiang, Y., and Li, H. Stable distance of persistent homology for dynamic graph comparison. Knowledge-Based Systems, 278: 0 110855, 2023

    2023

  50. [58]

    Simplex bounded confidence model for opinion fusion and evolution in higher-order interaction

    Ye, D., Lin, H., Jiang, H., Du, L., Li, H., Chen, Q., Wang, Y., and Yuan, L. Simplex bounded confidence model for opinion fusion and evolution in higher-order interaction. Expert Systems with Applications, pp.\ 126551, 2025

    2025

  51. [59]

    ROLAND : graph learning framework for dynamic graphs

    You, J., Du, T., and Leskovec, J. ROLAND : graph learning framework for dynamic graphs. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp.\ 2358--2366, 2022

    2022

  52. [60]

    Decoupled graph neural networks for large dynamic graphs

    Zheng, Y., Wei, Z., and Liu, J. Decoupled graph neural networks for large dynamic graphs. In Proceedings of the VLDB Endowment, volume 16, pp.\ 2239--2247, 2023

    2023

  53. [61]

    Meta- GNN : On few-shot node classification in graph meta-learning

    Zhou, F., Cao, C., Zhang, K., Trajcevski, G., Zhong, T., and Geng, J. Meta- GNN : On few-shot node classification in graph meta-learning. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management, pp.\ 2357--2360, 2019

    2019

  54. [62]

    Few-shot link prediction with meta-learning for temporal knowledge graphs

    Zhu, L., Xing, Y., Bai, L., and Chen, X. Few-shot link prediction with meta-learning for temporal knowledge graphs. Journal of Computational Design and Engineering, 10 0 (2): 0 711--721, 2023 a

    2023

  55. [63]

    WinGNN : Dynamic graph neural networks with random gradient aggregation window

    Zhu, Y., Cong, F., Zhang, D., Gong, W., Lin, Q., Feng, W., Dong, Y., and Tang, J. WinGNN : Dynamic graph neural networks with random gradient aggregation window. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp.\ 3650--3662, 2023 b

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.