Pith. sign in

REVIEW 5 major objections 6 minor 44 references

Understanding and Improving Laplacian Positional Encodings For Temporal GNNs

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Supra-Laplacian positional encodings for temporal graphs can be computed with fast iterative eigensolvers, and the approximate encodings match or slightly beat exact ones in downstream link prediction.

desk verdict Useful practical study of approximate eigensolvers for temporal Laplacian PEs, but the main recommendation is undersupported because approximate eigenvector accuracy is never measured. read the letter →

arxiv 2506.01596 v1 pith:MAVXEJFX submitted 2025-06-02 cs.LG cs.AI

classification cs.LGcs.AI
keywords temporalgraphneuralnetworkspositionalencodingssupra-LaplacianLaplacianeigenvectorsLOBPCGapproximateeigendecompositiondynamiclinkpredictionWeisfeiler-Lehmantest
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 tries to establish that positional encodings for temporal graph neural networks—vectors that tell a model where a node sits in an evolving structure—can be built from the low eigenvectors of the supra-Laplacian, a block matrix that couples each snapshot to its neighbors in time, and that these encodings survive replacing the exact eigendecomposition with a fast iterative approximation. If true, the main practical obstacle to supra-Laplacian encodings, their computational cost, is removed: the approximate LOBPCG variant runs up to 56 times faster on synthetic graphs with 50,000 active nodes while averaging 86.96% AUC across the tested models and datasets, marginally above the exact variant. The paper also gives a theoretical account, showing that these encodings minimize a sum of per-snapshot Laplacian smoothness plus an inter-snapshot consistency penalty, and proving that a temporal extension of the Weisfeiler-Lehman test is strictly stronger than applying the test to each snapshot separately. Empirically it maps where encodings help most: the gains are largest when node features carry little identifying information, and supra-Laplacian encodings generally beat per-snapshot Laplacian encodings, although the margin varies by model.

What carries the argument

The load-bearing object is the supra-Laplacian $L_{\text{supra}} = D_{\text{supra}} - A_{\text{supra}}$, a $T|V| \times T|V|$ block matrix whose diagonal blocks are snapshot adjacency matrices $A_t$ and whose off-diagonal blocks connect adjacent time slices, typically as $\mu I$. Its lowest eigenvectors become the positional encodings: node $v$ at time $t$ receives row $(t-1)|V|+v$ of the first $k$ eigenvectors. Two theoretical tools carry the argument. Proposition 1 rewrites the Rayleigh-quotient minimization of $L_{\text{supra}}$ as per-snapshot Laplacian quadratic forms plus an inter-layer consistency term, explaining why temporal encodings are smooth and sign-consistent across time. Proposition 2 introduces Supra-WL, a color-refinement scheme that hashes each node together with its temporal neighbors, and proves it strictly refines Layer-WL. On the computational side, the workhorse is LOBPCG, a block-preconditioned iterative eigensolver run for a small number of iterations; a trajectory variant concatenates intermediate iterates instead of keeping only the final approximation.

What would settle it

Run LOBPCG under a fixed small iteration budget on a temporal graph with more active nodes or a smaller spectral gap, compute the residual norm $\|L_{\text{supra}} v_i - \lambda_i v_i\|$ for each of the $k$ approximate eigenvectors, and check whether downstream link-prediction AUC tracks that residual. If residuals grow with graph size while AUC falls measurably below the exact-SLPE value, the claimed speedup-accuracy equivalence fails; if AUC stays flat despite growing residuals, the equivalence is genuinely a property of downstream insensitivity.

Watch

Extended reading notes

Core claim

The central discovery is that supra-Laplacian positional encodings for discrete-time dynamic graphs do not need exact eigenvectors to be useful. The authors prove that the lowest $k$ eigenvectors of the supra-Laplacian solve a smoothness trade-off: each time slice inherits the Laplacian structure of its own snapshot while a penalty term $\mu \sum_{t=2}^{T} \| X^{(t)} - X^{(t-1)} \|_F^2$ forces adjacent snapshots to have similar encodings. They then show that running a small, fixed number of iterations of the LOBPCG eigensolver produces encodings whose mean link-prediction AUC is comparable to, and in their aggregate slightly above, the exact variant (86.96% for SLPE-I versus 86.38% for SLPE-E), while cutting runtime by up to 56x. They also define Supra-WL, a temporal refinement of the Weisfeiler-Lehman test, and prove it is strictly more powerful than running WL on each snapshot independently, giving a formal sense in which the supra-adjacency carries distinctions no per-layer method can see.

Load-bearing premise

The load-bearing premise is that a small, fixed number of LOBPCG iterations produces eigenvectors whose downstream effect equals that of exact eigenvectors; the paper measures neither eigenvector residuals nor convergence, so the equivalence is inferred only from similar mean AUC on four small datasets.

Editorial extensions

If this is right

  • Practitioners can replace full eigendecompositions of the supra-Laplacian with LOBPCG approximations and expect essentially the same downstream AUC, which makes supra-Laplacian positional encodings practical for graphs with tens of thousands of active nodes.
  • Because supra-Laplacian encodings generally outperform per-snapshot Laplacian encodings in the reported settings, temporal connectivity is itself a useful signal for positional encoding, not just a computational complication.
  • Laplacian-based encodings give their largest lift when node features are least informative, so the encodings act partly as a substitute for node identity information.
  • Supra-WL being strictly stronger than Layer-WL implies that architectures that aggregate temporal neighbors can in principle distinguish temporal graphs that per-snapshot architectures cannot.
  • Exact solvers do not consistently beat approximate ones on the four datasets tested, so the cheaper solver is a reasonable default unless a specific dataset shows otherwise.

Reading between the lines

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

  • A natural extension, not tested in the paper, is an adaptive LOBPCG stopping rule: stop when the residual norm $\|L_{\text{supra}} v_i - \lambda_i v_i\|$ is small relative to the spectral gap, which would likely extend the reported speedup to denser and larger graphs.
  • The smoothness objective suggests that the inter-layer weight $\mu$, treated as fixed in the experiments, could be tuned per dataset; if SLPE quality tracks the optimal $\mu$, it would also give a principled way to choose the temporal window size.
  • Nothing in the theoretical argument requires discrete snapshots, so the same approximate-eigendecomposition recipe should transfer to continuous-time dynamic graphs by building supra-Laplacians over coarse time windows; the paper does not test this.
  • Because the paper reports only aggregated mean AUC without eigenvector residuals, the robustness of the speedup-accuracy trade-off to graph size and density remains open; a follow-up should plot iteration count versus both residual and downstream AUC.
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

5 major / 6 minor

Summary. The manuscript studies Laplacian positional encodings for temporal GNNs. It first formalizes supra-Laplacian PEs (SLPEs) as the minimizers of an objective that balances intra-snapshot Laplacian smoothness with inter-snapshot consistency (Prop. 1), and it introduces a Supra-WL test to argue that supra-Laplacian representations are strictly more expressive than per-snapshot WL colorings (Prop. 2). The paper then proposes to compute SLPEs with iterative eigensolvers—Lanczos for exact and LOBPCG for inexact computation—and a trajectory-based variant that concatenates intermediate solver iterates. The empirical section evaluates four temporal GNN architectures (EGCN, GRUGCN, HTGN, SLATE) on four discrete-time datasets (CanParl, as733, dblp, enron10) under three node-feature settings (one-hot, random, constant zero), measuring dynamic link prediction AUC. The authors report that PEs help in about 70% of cases, that SLPEs generally outperform LPEs, that approximate (I) variants are close to exact (E) variants, and that LOBPCG is up to 56x faster than Lanczos on large synthetic graphs. They conclude that SLPE-I is a robust default.

Significance. If the claims hold, the paper would make a useful contribution: it gives a theoretical justification for supra-Laplacian encodings, a concrete efficiency recipe for practitioners, and a broad empirical map of when PEs help across architectures and feature regimes. The algebraic proof of Proposition 1 is correct, and the empirical study is unusually broad for this area (4 models × 4 datasets × 3 feature schemes). The main weakness is that the practical recommendation 'SLPE-I as robust default' rests on average AUC differences that are not statistically supported, and on a timing comparison whose approximation accuracy is never quantified. With additional convergence diagnostics and significance testing, the practical claims would be substantially strengthened.

major comments (5)
  1. [§6.2, Tables 5–6] The claim that SLPE-I is a 'robust default' is not supported by the reported statistics. The headline gap between SLPE-I (86.96%) and SLPE-E (86.38%) in Table 5 is smaller than the dispersion reported in Table 6, where the E−I difference quartiles include negative values for every model (e.g., EGCN [Q1,Q3]=[−0.68,0.57], SLATE [−0.32,0.36]). No significance tests, confidence intervals, or paired comparisons are reported, and the mean differences are computed over a small number of model–dataset–feature cells. Please add per-cell significance tests or at least bootstrap CIs and report how many individual comparisons favor SLPE-I over SLPE-E.
  2. [§5 and §6.2, Appendix D.1] The paper never quantifies the accuracy of the approximate eigenvectors used by LOBPCG. The comparison between inexact (I) and exact (E) variants is only in downstream AUC on four small datasets, and the timing comparison is a fixed-budget comparison: LOBPCG is run with maxiter ∈ {5,10,20,50} (Appendix D.1) while Lanczos is run 'until convergence' (Section 6.1). No eigenvector residual norms, no subspace distances to the exact eigenspace, and no AUC-versus-iteration curves are reported. Because the central recommendation is that approximate SLPEs retain most of the benefit of exact SLPEs, please provide convergence diagnostics and an accuracy–runtime tradeoff curve (e.g., AUC or residual vs. LOBPCG iterations) on at least the larger datasets.
  3. [§6.2, Figure 6] The 56x speedup and the 50,000-node scalability claim are demonstrated on static Barabási–Albert graphs, not on supra-Laplacians built from temporal snapshots. The abstract and Section 6.2 state these results in the context of SLPE computation, but Figure 6 times the eigendecomposition of a single static Laplacian. The runtime advantage of LOBPCG on a static graph does not automatically transfer to the supra-Laplacian of a temporal graph sequence, whose block structure and conditioning differ. Please either measure the supra-Laplacian case directly or clearly qualify the claim.
  4. [§4.1, Eq. (4) and Appendix C.1] Proposition 1 is proved for a weighted supra-Laplacian with an inter-layer parameter µ, but the definition in Section 3, Eq. (2), sets the inter-layer blocks B_ij to the identity matrix with no µ, and µ does not appear among the hyperparameters in Appendix D.1. The proof in C.1 also uses a specific degree correction (D_t + µI or D_t + 2µI) that is not stated in the main text. Please clarify whether the implemented SLPEs correspond to µ=1, and if so state that explicitly, or add µ to the hyperparameter description; otherwise the theoretical smoothness result is not connected to the empirical SLPEs.
  5. [§4.2, Fig. 4 and Appendix C.2] The strictness part of Proposition 2 rests on the pair of graphs in Figure 4, but the figure is not accompanied by a formal description of the two DTDGs (node sets, edge sets per snapshot, or the coloring that distinguishes them). The proof in Appendix C.2 establishes that Supra-WL refines Layer-WL, which is only half of the strictness claim. Please provide a precise specification of the example and the color refinement sequence, so the strictness can be checked without relying on the drawing.
minor comments (6)
  1. [Figures 5–6] The runtime plots report single timing measurements without error bars or the number of repeats; since the central speedup claims are based on these plots, please report mean ± std over at least three runs.
  2. [§6.1] The statement that Lanczos is run 'till convergence' does not specify a convergence tolerance; please state the criterion so the timing comparison is reproducible.
  3. [§3, Eq. (2)] In Eq. (2), B_ij denotes n×n blocks, while the text says 'Bij to be the identity matrix I when |i−j|=1'; please clarify that all non-adjacent blocks are zero and that the identity blocks are n×n.
  4. [§5] The trajectory-based sign handling randomly chooses a sign per eigenvector; please state whether the same random sign is used across the five runs, since this randomization can inflate variance in the reported AUC tables.
  5. [Appendix D.2, Table 7] Table 7 uses the heading 'randn' while the text and Tables 3–4 use 'random'; please unify the terminology.
  6. [Appendix A.1] The split table lists 'Colab' as a dataset, but Colab is not among the four datasets used in the experiments (CanParl, as733, dblp, enron10); this appears to be a leftover and should be fixed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the theoretical propositions are derived from definitions, and the empirical claims are measured rather than fitted by construction.

full rationale

The paper's central theoretical and empirical claims do not reduce to their own inputs. Proposition 1 establishes an equivalence between supra-Laplacian eigenvectors and a smoothness objective; this is a standard Rayleigh-quotient characterization derived from the block structure of the supra-Laplacian, not a self-referential definition. Proposition 2 is proven independently in Appendix C.2 via an induction showing Supra-WL refines Layer-WL, together with an explicit non-isomorphic example. The computational claims are timing and accuracy measurements on test snapshots: LOBPCG and Lanczos are standard algorithms, and the reported AUC differences are computed on held-out data rather than constructed from the fitted parameters. The trajectory-based approach is inspired by the same group's prior work [6], but the paper evaluates trajectory variants empirically and does not rely on [6] to justify its conclusions; the self-citation is a building block, not a load-bearing circular justification. The lack of eigenvector-residual or convergence checks in the inexact-solver comparison is a legitimate concern about evidence quality and transferability, but it is a missing-controls issue, not a circularity. Therefore no circular step can be exhibited, and the appropriate score is 0.

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

The central empirical claims rest on hyperparameters (window size, number of eigenvectors, iteration counts) tuned on validation data, and on assumptions about the supra-Laplacian construction and the evaluation protocol that are inherited from prior work without independent validation. No genuinely new physical or conceptual entities are introduced by this paper.

free parameters (5)
  • window size ws = integer sampled from [2,5]
    Number of snapshots used to build the supra-Laplacian; tuned by Bayesian search per dataset (Appendix D.1). Central to the SLP definition.
  • number of eigenvectors k = integer sampled from [4,16]
    Dimensionality of the positional encoding; tuned per dataset via Bayesian search. Directly affects model input size and performance.
  • max iterations for iterative solvers = 5, 10, 20, 50
    Stopping point for LOBPCG and trajectory generation; influences the speed/accuracy tradeoff.
  • inter-layer weight mu = not reported (assumed 1)
    Controls inter-layer coupling in Proposition 1, but Section 3 defines B_ij = I without specifying the actual weight used in experiments.
  • PE initialization = normal, rademacher, uniform, with_old_pes
    Chosen during hyperparameter search; part of the PE computation pipeline.
assumptions (5)
  • standard math Spectral theorem for symmetric matrices and trace-minimization characterization of eigenvectors
    Used in Proposition 1 and proof in Appendix C.1 to equate supra-Laplacian eigenvectors with minimizers of tr(X^T L X).
  • domain assumption Supra-Laplacian block structure with inter-layer identity coupling
    Section 3 assumes temporal dependencies are captured by identity matrices between consecutive snapshots; this is a modeling choice, not a derived fact.
  • standard math 1-WL color refinement and HASH injectivity
    Used in Proposition 2 and Appendix C.2; the proofs rely on HASH being injective.
  • domain assumption Evaluation protocol (negative sampling, windowing, dataset splits) is representative
    The paper follows prior work [43,41] for splits, but does not specify negative sampling details, so the AUC numbers are conditional on an unstated protocol.
  • ad hoc to paper Eigenvector sign ambiguity can be resolved by random sign choice
    Section 5 and [5] use random sign flips; the trajectory method assumes this yields valid encodings, which is an empirical assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Understanding and Improving Laplacian Positional Encodings For Temporal GNNs." pith.science (2026). https://pith.science/paper/MAVXEJFX

@misc{pith2026250601596,
  author       = {Pith},
  title        = {Pith review of: Understanding and Improving Laplacian Positional Encodings For Temporal GNNs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MAVXEJFX}},
  note         = {Machine review of arXiv:2506.01596}
}
read the original abstract

Temporal graph learning has applications in recommendation systems, traffic forecasting, and social network analysis. Although multiple architectures have been introduced, progress in positional encoding for temporal graphs remains limited. Extending static Laplacian eigenvector approaches to temporal graphs through the supra-Laplacian has shown promise, but also poses key challenges: high eigendecomposition costs, limited theoretical understanding, and ambiguity about when and how to apply these encodings. In this paper, we address these issues by (1) offering a theoretical framework that connects supra-Laplacian encodings to per-time-slice encodings, highlighting the benefits of leveraging additional temporal connectivity, (2) introducing novel methods to reduce the computational overhead, achieving up to 56x faster runtimes while scaling to graphs with 50,000 active nodes, and (3) conducting an extensive experimental study to identify which models, tasks, and datasets benefit most from these encodings. Our findings reveal that while positional encodings can significantly boost performance in certain scenarios, their effectiveness varies across different models.

Figures

Figures reproduced from arXiv: 2506.01596 by the authors.

Figure 1
Figure 1. An overview of our proposed fast SLPEs computation procedure. The [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Examples of adjacency matrices. Left: an adjacency matrix of a single [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Comparison of eigenvector smoothness in multilayer single path graphs. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Supra-WL correctly distinguishes the depicted non-isomorphic temporal [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Time (ms) performance comparison of Full Eigendecomposition, Lanczos, [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Time (ms) comparison of Full Eigendecomposition, Lanczos, and [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 35 canonical work pages

  1. [1]

    In: Proceedings of the Thirtieth International Joint Conference on Artifical Intelligence (IJCAI) (2021)

    Abboud,R.,Ceylan,İ.İ.,Grohe,M.,Lukasiewicz,T.:Thesurprisingpowerofgraph neural networks with random node initialization. In: Proceedings of the Thirtieth International Joint Conference on Artifical Intelligence (IJCAI) (2021)

  2. [2]

    Neural Comput.15(6), 1373–1396 (Jun 2003)

    Belkin, M., Niyogi, P.: Laplacian eigenmaps for dimensionality reduction and data representation. Neural Comput.15(6), 1373–1396 (Jun 2003)

  3. [3]

    Cong, W., Zhang, S., Kang, J., Yuan, B., Wu, H., Zhou, X., Tong, H., Mahdavi, M.: Do we really need complicated model architectures for temporal networks? In: The Eleventh International Conference on Learning Representations (2023), https://openreview.net/forum?id=ayPPc0SyLv1

  4. [4]

    In: In- ternational Conference on Learning Representations (2021),https://openreview

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: In- ternational Conference on Learning Representations (2021),https://openreview. net/forum?id=YicbFdNTTy

  5. [5]

    Journal of Machine Learning Research24(43), 1–48 (2023)

    Dwivedi, V.P., Joshi, C.K., Luu, A.T., Laurent, T., Bengio, Y., Bresson, X.: Bench- marking graph neural networks. Journal of Machine Learning Research24(43), 1–48 (2023)

  6. [6]

    In: Proceedings of the 40th International Conference on Machine Learning

    Eliasof, M., Frasca, F., Bevilacqua, B., Treister, E., Chechik, G., Maron, H.: Graph positional encoding via random feature propagation. In: Proceedings of the 40th International Conference on Machine Learning. ICML’23, JMLR.org (2023)

  7. [7]

    In: 2023 IEEE Ninth Multimedia Big Data (BigMM)

    Fard, S.H., Ghassemi, M.: Temporal Link Prediction Using Graph Embedding Dynamics . In: 2023 IEEE Ninth Multimedia Big Data (BigMM). pp. 48–55. IEEE Computer Society, Los Alamitos, CA, USA (Dec 2023).https://doi.org/ 10.1109/BigMM59094.2023.00014, https://doi.ieeecomputersociety.org/10. 1109/BigMM59094.2023.00014

  8. [8]

    Fey, M., Lenssen, J.E.: Fast graph representation learning with pytorch geometric (2019), https://arxiv.org/abs/1903.02428

Show all 44 references
  1. [9]

    Gómez, S., Díaz-Guilera, A., Gómez-Gardeñes, J., Pérez-Vicente, C.J., Moreno, Y., Arenas, A.: Diffusion dynamics on multiplex networks. Phys. Rev. Lett.110(2), 028701 (Jan 2013)

  2. [10]

    Heo, B., Park, S., Han, D., Yun, S.: Rotary position embedding for vision trans- former.In:EuropeanConferenceonComputerVision.pp.289–305.Springer(2024)

  3. [11]

    Advances in Neural Information Processing Systems 36, 2056–2073 (2023)

    Huang, S., Poursafaei, F., Danovitch, J., Fey, M., Hu, W., Rossi, E., Leskovec, J., Bronstein, M., Rabusseau, G., Rabbany, R.: Temporal graph benchmark for machine learning on temporal graphs. Advances in Neural Information Processing Systems 36, 2056–2073 (2023)

  4. [12]

    In: The Third Learning on Graphs Conference (2024),https://openreview.net/forum? id=ZKHV6Cpsxg

    Huang, S., Poursafaei, F., Rabbany, R., Rabusseau, G., Rossi, E.: UTG: Towards a unified view of snapshot and event based models for temporal graphs. In: The Third Learning on Graphs Conference (2024),https://openreview.net/forum? id=ZKHV6Cpsxg

  5. [13]

    Huang, Y., Lu, W., Robinson, J., Yang, Y., Zhang, M., Jegelka, S., Li, P.: On the stability of expressive positional encodings for graphs. In: The Twelfth Interna- tional Conference on Learning Representations (2024),https://openreview.net/ forum?id=xAqcJ9XoTf Understanding La...

  6. [14]

    In: The Thirty-eighth Annual Con- ference on Neural Information Processing Systems (2024),https://openreview

    Karmim, Y., Lafon, M., Fournier-S’niehotta, R., THOME, N.: Supra-laplacian en- coding for transformer on dynamic graphs. In: The Thirty-eighth Annual Con- ference on Neural Information Processing Systems (2024),https://openreview. net/forum?id=vP9qAzr2Gw

  7. [15]

    Kazemi, S.M., Goel, R., Jain, K., Kobyzev, I., Sethi, A., Forsyth, P., Poupart, P.: Representation learning for dynamic graphs: a survey. J. Mach. Learn. Res.21(1) (Jan 2020)

  8. [16]

    Knyazev, A.: Recent implementations, applications, and extensions of the locally optimal block preconditioned conjugate gradient method (lobpcg) (2017),https: //arxiv.org/abs/1708.08354

  9. [17]

    KDD2019, 1269–1278 (Aug 2019)

    Kumar, S., Zhang, X., Leskovec, J.: Predicting dynamic embedding trajectory in temporal interaction networks. KDD2019, 1269–1278 (Aug 2019)

  10. [18]

    Kuncheva, Z., Kounchev, O.: Spectral properties of the laplacian of tempo- ral networks following a constant block jacobi model. Phys. Rev. E 109, 064309 (Jun 2024). https://doi.org/10.1103/PhysRevE.109.064309, https:// link.aps.org/doi/10.1103/PhysRevE.109.064309

  11. [19]

    Lanczos, C.: An iteration method for the solution of the eigenvalue problem of linear differential and integral operators. J. Res. Natl. Bur. Stand. B45, 255–282 (1950). https://doi.org/10.6028/jres.045.026

  12. [20]

    In: The Eleventh International Conference on Learning Representations (2023), https: //openreview.net/forum?id=Q-UHqMorzil

    Lim, D., Robinson, J.D., Zhao, L., Smidt, T., Sra, S., Maron, H., Jegelka, S.: Sign and basis invariant networks for spectral graph representation learning. In: The Eleventh International Conference on Learning Representations (2023), https: //openreview.net/forum?id=Q-UHqMorzil

  13. [21]

    In: Science of Cyber Security, pp

    Lin, W., Zhou, S., Li, M., Chen, G.: Dismantling interdependent networks based on supra-laplacian energy. In: Science of Cyber Security, pp. 205–213. Lecture notes in computer science, Springer International Publishing, Cham (2021)

  14. [22]

    In: NeurIPS 2022 Workshop on Symmetry and Geometry in Neural Representations (2022),https: //openreview.net/forum?id=BNhhZwAlVNC

    Maskey, S., Parviz, A., Thiessen, M., Stärk, H., Sadikaj, Y., Maron, H.: Generalized laplacian positional encoding for graph representation learning. In: NeurIPS 2022 Workshop on Symmetry and Geometry in Neural Representations (2022),https: //openreview.net/forum?id=BNhhZwAlVNC

  15. [23]

    Transactions on Machine Learning Research (2024), https: //openreview.net/forum?id=BTgHh0gSSc

    Ozmen, M., Markovich, T.: Recent link classification on temporal graphs us- ing graph profiler. Transactions on Machine Learning Research (2024), https: //openreview.net/forum?id=BTgHh0gSSc

  16. [24]

    In: Proceedings of the AAAI conference on artificial intelli- gence

    Pareja, A., Domeniconi, G., Chen, J., Ma, T., Suzumura, T., Kanezashi, H., Kaler, T., Schardl, T., Leiserson, C.: Evolvegcn: Evolving graph convolutional networks for dynamic graphs. In: Proceedings of the AAAI conference on artificial intelli- gence. vol. 34, pp. 5363–5370 (2020)

  17. [25]

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Köpf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: Pytorch: An imperativ...

  18. [26]

    ACM Comput

    Qin, M., Yeung, D.Y.: Temporal link prediction: A unified framework, taxonomy, and review. ACM Comput. Surv.56(4) (Nov 2023). https://doi.org/10.1145/ 3625820, https://doi.org/10.1145/3625820

  19. [27]

    Advances in Neural Infor- mation Processing Systems35 (2022)

    Rampášek, L., Galkin, M., Dwivedi, V.P., Luu, A.T., Wolf, G., Beaini, D.: Recipe for a General, Powerful, Scalable Graph Transformer. Advances in Neural Infor- mation Processing Systems35 (2022)

  20. [28]

    In: ICML 2020 Work- shop on Graph Representation Learning (2020) 16 Y

    Rossi, E., Chamberlain, B., Frasca, F., Eynard, D., Monti, F., Bronstein, M.: Tem- poral graph networks for deep learning on dynamic graphs. In: ICML 2020 Work- shop on Graph Representation Learning (2020) 16 Y. Galron et al

  21. [29]

    arXiv [physics.soc-ph] (Sep 2019)

    Sato, K., Oka, M., Barrat, A., Cattuto, C.: DyANE: Dynamics-aware node embed- ding for temporal networks. arXiv [physics.soc-ph] (Sep 2019)

  22. [30]

    net/forum?id=S19eAF9ee

    Seo, Y., Defferrard, M., Vandergheynst, P., Bresson, X.: Structured sequence mod- eling with graph convolutional recurrent networks (2017),https://openreview. net/forum?id=S19eAF9ee

  23. [31]

    https://doi.org/10.1109/ACCESS.2021.3082932

    Skarding, J., Gabrys, B., Musial, K.: Foundations and modeling of dynamic net- worksusingdynamicgraphneuralnetworks:Asurvey.IEEEAccess 9,79143–79168 (2021). https://doi.org/10.1109/ACCESS.2021.3082932

  24. [32]

    Solé-Ribalta, A., De Domenico, M., Kouvaris, N.E., Díaz-Guilera, A., Gómez, S., Arenas, A.: Spectral properties of the laplacian of multiplex networks. Phys. Rev. E 88, 032807 (Sep 2013).https://doi.org/10.1103/PhysRevE.88.032807, https: //link.aps.org/doi/10.1103/PhysRevE.88.032807

  25. [33]

    Souza, A.H., Mesquita, D., Kaski, S., Garg, V.K.: Provably expressive temporal graphnetworks.In:Oh,A.H.,Agarwal,A.,Belgrave,D.,Cho,K.(eds.)Advancesin Neural Information Processing Systems (2022),https://openreview.net/forum? id=MwSXgQSxL5s

  26. [34]

    In: 2022 IEEE International Conference on Big Data (Big Data)

    Sun, J., Gu, M., Yeh, C.C.M., Fan, Y., Chowdhary, G., Zhang, W.: Dynamic graph node classification via time augmentation. In: 2022 IEEE International Conference on Big Data (Big Data). pp. 800–805. IEEE (2022)

  27. [35]

    In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=HyePrhR5KX

    Trivedi, R., Farajtabar, M., Biswal, P., Zha, H.: Dyrep: Learning representations over dynamic graphs. In: International Conference on Learning Representations (2019), https://openreview.net/forum?id=HyePrhR5KX

  28. [36]

    Advances in neural information pro- cessing systems 30 (2017)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems 30 (2017)

  29. [37]

    In: International Con- ference on Learning Representations (2021),https://openreview.net/forum?id= KYPz4YsCPj

    Wang, Y., Chang, Y.Y., Liu, Y., Leskovec, J., Li, P.: Inductive representation learning in temporal networks via causal anonymous walks. In: International Con- ference on Learning Representations (2021),https://openreview.net/forum?id= KYPz4YsCPj

  30. [38]

    In: Proceedings of the Eighteenth ACM International Conference on Web Search and Data Mining (2025)

    Wang, Z., Zhou, S., Chen, J., Zhang, Z., Hu, B., Feng, Y., Chen, C., Wang, C.: Dynamic graph transformer with correlated spatial-temporal positional encoding. In: Proceedings of the Eighteenth ACM International Conference on Web Search and Data Mining (2025)

  31. [39]

    In: International Conference on Learning Representations (ICLR) (2020)

    da Xu, chuanwei ruan, evren korpeoglu, sushant kumar, kannan achan: Induc- tive representation learning on temporal graphs. In: International Conference on Learning Representations (ICLR) (2020)

  32. [40]

    IEEE Access12, 43460–43484 (2024)

    Yang, L., Chatelain, C., Adam, S.: Dynamic graph representation learning with neural networks: A survey. IEEE Access12, 43460–43484 (2024)

  33. [41]

    In: Pro- ceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining

    Yang, M., Zhou, M., Kalander, M., Huang, Z., King, I.: Discrete-time temporal network embedding via implicit hierarchical learning in hyperbolic space. In: Pro- ceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. pp. 1975–1985 (2021)

  34. [42]

    Yang, Y., Tu, L., Guo, T., Chen, J.: Spectral properties of supra-laplacian for par- tially interdependent networks. Appl. Math. Comput.365(124740), 124740 (Jan 2020)

  35. [43]

    Advances in Neural Information Processing Systems 36, 67686–67700 (2023)

    Yu, L., Sun, L., Du, B., Lv, W.: Towards better dynamic graph learning: New ar- chitecture and unified library. Advances in Neural Information Processing Systems 36, 67686–67700 (2023)

  36. [44]

    yes" to a bill. The line’s thickness shows how often one MP supported another with

    Zhang, X., Wang, Y., Wang, X., Zhang, M.: Efficient neural common neighbor for temporal graph link prediction (2024),https://arxiv.org/abs/2406.07926 Supplementary Material: Understanding and Improving Laplacian Positional Encodings For Temporal GNNs A Datasets Datasets statis...

Pith tools

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