Pith. sign in

REVIEW 5 major objections 5 minor 57 references

Learnable Spatial-Temporal Positional Encoding for Link Prediction

T0 review · 5 major / 5 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 L-STEP learns node positions from past snapshots with a learnable Fourier filter, proving MLPs can match attention-based transformers on temporal link prediction.

desk verdict L-STEP's empirical contribution is solid and worth engaging with, but its advertised spectral-preservation theorem is not supported by the proof in Appendix C, and the abstract overstates the TGB results. read the letter →

arxiv 2506.08309 v2 pith:NODXWMTB submitted 2025-06-10 cs.LG cs.AI

Katherine Tieu , Dongqi Fu , Zihao Li , Ross Maciejewski , Jingrui He This is my paper
classification cs.LGcs.AI
keywords temporallinkpredictionpositionalencodinglearnablediscreteFouriertransformspectralgraphtheoryneuralnetworksgraphs
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a learnable, time-evolving positional encoding can carry enough spatial-temporal information for temporal link prediction that a simple MLP-only model matches or beats transformer-based state of the art. The proposed model, L-STEP, approximates a node's current positional encoding by applying a Discrete Fourier Transform to its recent past encodings, filtering in the frequency domain with a learned complex filter, and pooling the result with learned weights. The paper proves a spectral-preservation bound (Theorem 4.1): on graphs whose Laplacian positional encodings change slowly, the approximate encoding stays close to the encoding of future snapshots, so the model does not need to see the future topology to represent it. Empirically, L-STEP reports the best average rank against ten algorithms on thirteen classic datasets under several sampling strategies, and competitive results on the large-scale TGB benchmark, with complexity linear in the number of nodes rather than edges. The central bet is that learnable spectral filtering of past positions is a cheaper substitute for attention.

What carries the argument

The Learnable Positional Encoding (LPE) module: it treats a node's past positional encodings as an $L$-length time series, applies the Discrete Fourier Transform, multiplies by a learnable complex-valued filter $W_{\text{filter}} \in \mathbb{C}^{d_P \times L}$ to suppress noise in the frequency domain, transforms back with the inverse DFT, and sum-pools with learnable weights to produce $\hat{p}_u^t$. The Node-Link-Positional Encoder then combines these with node and edge features via MLPs only. The spectral argument rests on the fact that the Fourier basis diagonalizes the ring-graph Laplacian (Lemma C.1), which lets the proof rewrite the filtering operation as an application of the ring Laplacian $L(R_L)$ followed by $\Lambda W_{\text{filter}}$; the bound then follows from the slow-change assumption plus the imposed filter bound $\sum_i |(\Lambda W_{\text{filter}})_i| < 1$ and normalization $\sum_j \alpha_j = 1$.

What would settle it

Train L-STEP on a temporal graph whose Laplacian positional encodings change abruptly between consecutive snapshots (for example, a graph where central hubs appear and vanish at every timestamp), and record $\|\hat{p}_u^t - \hat{p}_u^{t''}\|_2$ along with test AP; Theorem 4.1 provides no bound in this regime, and the model's edge over DyGFormer should shrink or disappear. A second check: during training, compute $\sum_i |(\Lambda W_{\text{filter}})_i|$ on validation snapshots; whenever this quantity exceeds 1, the derived bound in Eq. 21 no longer holds even on slowly changing graphs, so the spectral-preservation claim is directly testable.

Watch

Extended reading notes

Core claim

L-STEP's central claim is that positional encoding for temporal graphs can be learned rather than fixed: given the sequence of a node's positional encodings over the L most recent timestamps, applying the DFT, multiplying by a learnable complex filter $W_{\text{filter}}$ in the frequency domain, applying the inverse DFT, and pooling with learnable weights yields an approximate encoding $\hat{p}_u^t$ that preserves the graph's spectral structure. Theorem 4.1 states that if the ground-truth Laplacian positional encodings are approximately constant across snapshots ($p^0 \approx \cdots \approx p^{t-1}$), then $\|\hat{p}_u^t - \hat{p}_u^{t''}\|_2 \le \varepsilon$ with $\varepsilon$ depending on $L$, which the paper interprets as showing that the learned encoding represents future positional information without peeking at future topology. The paper also claims that MLPs, not attention, suffice to exploit this encoding, and that L-STEP achieves the best average rank on 13 temporal link prediction datasets in transductive and inductive settings while having $O(n)$ complexity.

Load-bearing premise

The guarantee that the learned encoding carries future positional information holds only when the graph's Laplacian positional encodings stay nearly unchanged across consecutive snapshots (the 'slowly changing' assumption), and the proof additionally requires the trained filter's weighted spectrum to stay below 1 and the pooling weights to sum to 1; the theorem also bounds the gap between two approximate encodings, not the gap between an approximate and the true encoding.

Editorial extensions

If this is right

  • If Theorem 4.1 holds, L-STEP can produce useful positional encodings for future snapshots without recomputing a Laplacian eigendecomposition at each timestamp, avoiding $O(|V|^3)$ or $O(|E|^{3/2})$ per-step costs.
  • The $O(n(L\log L + K))$ per-step complexity with constants $L$, $K$, and $t_{\text{gap}}$ independent of $n$ shows that node-based, rather than edge-pair-based, aggregation can suffice for temporal link prediction, contrasting with the co-occurrence design of DyGFormer and FreeDyG.
  • On the 13 classic datasets, the reported average ranks (e.g., 1.85 for transductive AP) imply that an MLP-only predictor on this encoding is competitive or better than attention-based models, so attention may not be the load-bearing component in temporal link prediction performance.
  • The robustness test replacing Laplacian PE initialization with Random Walk PE indicates that the LPE scheme is not tied to any specific initial positional encoding.
  • On TGB, the 2nd and 3rd place ranks on tgbl-review and tgbl-coin suggest the approach transfers to graphs with millions of edges.

Reading between the lines

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

  • The proof's slow-change condition is essentially a smoothness assumption on graph spectra; a natural testable extension is to measure, on each dataset, the drift in Laplacian positional encodings across snapshots and check whether L-STEP's margin over baselines correlates with that drift.
  • If the claim that MLPs fully exploit the encoding is right, the same LPE scheme could replace attention in other temporal graph tasks beyond link prediction, such as node classification or edge regression on continuous-time graphs, at reduced cost.
  • The bound in Theorem 4.1 concerns the distance between two approximate encodings at different times, not the error with respect to the true current encoding; an honest check would compute $\|\hat{p}_u^t - p_u^t\|_2$ on real data and compare it to the derived $\varepsilon$ bound.
  • The learnable filter imposes a hand-chosen spectral radius condition ($\sum_i |(\Lambda W_{\text{filter}})_i| < 1$) that is not enforced during training; monitoring this quantity during training would show whether the theorem's regime is actually entered.
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 / 5 minor

Summary. The paper proposes L-STEP, a temporal link prediction model built from a Learnable Positional Encoding module (LPE) and a Node-Link-Positional Encoder, both using MLPs rather than attention. LPE applies a learnable Fourier-domain filter to a length-L history of per-node positional encodings, then pools the filtered sequence to form an approximate encoding for the current time; stored encodings are updated after ground-truth links are revealed. The paper claims a spectral-preservation theorem (Theorem 4.1) saying that under a "slowly changing" temporal graph the approximate encoding remains close to its future values without using future topology, an O(n) complexity analysis, and strong empirical results on 13 classic datasets under transductive and inductive settings with three negative sampling strategies, plus competitive results on two large TGB datasets.

Significance. If the empirical results are taken at face value, the contribution is practically significant: a lightweight MLP-only model with per-step complexity O(n(Llog L+K+tgap)) that reaches or beats attention-based state-of-the-art on the benchmark suite is a useful data point for the field. The experimental protocol is solid, including chronological 70/15/15 splits, five seeds, validation-based early stopping, three negative sampling strategies, ablation studies, and runtime comparisons. However, the theoretical guarantee advertised in the abstract and Section 4 is not established by the proof in Appendix C, so the paper's significance currently rests on the experiments rather than on the spectral-preservation claim. The theoretical claim needs substantial repair before the paper can be accepted in its advertised form.

major comments (5)
  1. [Appendix C, Eq. (16)] The proof's first displayed identity substitutes p^{L+1} with p^1 in the last coordinate, turning p_L - p_{L+1} + p_L - p_{L-1} into p_L - p_1 + p_L - p_{L-1}. This is an additional future-stationarity assumption p^{L+1} ≈ p^1, which is not entailed by the theorem premise p^0 ≈ ... ≈ p^{t-1} and cannot be checked without looking at future topology. Since the advertised no-future-topology guarantee depends on this step, Theorem 4.1 is not established as stated.
  2. [Appendix C, proof of Theorem 4.1] The proof assumes p^0, ..., p^L are Laplacian eigenvectors of their respective graph snapshots, but in the implemented model the stored encodings are produced by Eq. (3), an MLP update with tanh, after initializing with Laplacian eigenvectors. No argument shows that these learned vectors remain close to ground-truth eigenvectors over time, so a bound on eigenvector sequences does not transfer to the sequence actually stored and used by L-STEP.
  3. [Section 4, Theorem 4.1 and Appendix C] Even if the inequality ||ep^{L+1} - ep^{L+2}|| were proved correctly, the theorem's advertised conclusion is that the encoding "can be a good representation of the positional information of future graph snapshots." The derived bound compares two approximate encodings, not an approximate encoding with the true positional encoding of the future snapshot. This gap between the theorem's statement and its proof is not addressed in Appendix C.
  4. [Appendix C, Lemma C.1 and Eq. (21)] Lemma C.1 states the ring-graph Laplacian eigenvalues as 2 - cos(2πk/n); the correct eigenvalues for the unnormalized ring Laplacian are 2 - 2cos(2πk/n). This error propagates into the diagonal matrix Λ used in Eqs. (17)-(18) and into the filter contraction condition. In addition, the derivation of Eq. (21) silently imposes the normalization sum α_j = 1 and the filter condition sum_i |(ΛW_filter)_i| < 1 without listing them as assumptions in Theorem 4.1, and the factor (2L-2) is asserted rather than derived from the preceding O(L^2) coefficient sum.
  5. [Section 4, Theorem 4.1] The theorem's premise p^0 ≈ ... ≈ p^{t-1} asserts approximate constancy of ground-truth positional encodings over time, while the conclusion is about approximate encodings; the proof additionally needs p^{L+1} ≈ p^1. The relation between the premise and the conclusion is therefore close to assuming the desired stability behavior rather than deriving it from graph dynamics. At minimum, the theorem must separate ground-truth eigenvalue stability from approximation-error stability, and the proof must state which of these is actually used at each step.
minor comments (5)
  1. [Section 4 and Appendix C] The notation in Theorem 4.1 uses t and t'', while the proof switches to ep^{L+1} and ep^{L+2}; please align these notations and define how t'' relates to L and the indexing of snapshots.
  2. [Eq. (20) and Eq. (21)] The indices for the weights α are inconsistent: the text says α_1, ..., α_{t-1} in one place and α_1, ..., α_L in the following display. Please define α once and use consistent indexing throughout the proof.
  3. [Appendix H.5.5] The subsection title says "parameter analysis for α_neg" but the table reports results for α_pe; the title and the accompanying text should be corrected.
  4. [Abstract and Section 5.1] The abstract says 10 algorithms, while Section 5.1 lists 9 baseline methods plus L-STEP; please state the count consistently so that the comparison is unambiguous.
  5. [Section 3.1 and Algorithm 3] The text says LPE "does not peek" at current topology, but Algorithm 3 reveals ground-truth links after prediction and updates p^t via Eq. (3). Please clarify that the no-peek property applies only to the forward computation of ep^t, not to the training-time update of stored encodings.

Circularity Check

1 steps flagged · score 6.0 of 10

Theorem 4.1's future-representation guarantee is derived by silently substituting p^{L+1}=p^1, an unstated future-stationarity premise; the proof therefore reduces to the target behavior, while the empirical benchmark evaluation remains independent.

  1. other [Appendix C, proof of Theorem 4.1, Eq. 16 (and Section 4, Theorem 4.1)]
    "epL+1 − epL+2 + epL+1 − epL + ( [p0 −pL 0 0. . .pL+1 −p1] ) F W_filter F^{-1} [α1 . . . α_{t−1}] = ... = ( [p1 −p2 +p1 −pL . . .p_i −p_{i+1} +p_i −p_{i−1} . . .p_L −p1 +p_L −p_{L−1}] ) F W_filter F^{-1} [α1 . . . αL]"

    The proof identifies ep^{L+1} with the approximate encoding at time t and ep^{L+2} with the approximate encoding at the future time t''. In the displayed equation, the last coordinate p_L−p_{L+1}+p_L−p_{L−1} is rewritten as p_L−p_1+p_L−p_{L−1}, which is exactly the substitution p^{L+1}=p^1. But p^{L+1} is the ground-truth positional encoding of the current/future snapshot, and Theorem 4.1's premise p0≈...≈p^{t−1} stops before p^{L+1}; the proof never derives p^{L+1}≈p^1 without peeking at future topology. Thus the advertised conclusion — that L-STEP's encoding represents future positional information without future graph topology — is obtained by assuming the very current/future ground-truth stability that the theorem claims to establish.

full rationale

The empirical contribution of L-STEP is not circular: performance is measured against external baselines and public benchmarks (TGB), the implementation is released, and the MLP-versus-transformer and ablation studies compare against the paper's own components rather than re-fitting the headline numbers. Self-citations appear in the introduction and related work but are contextual, not load-bearing for the method's claims, so they do not raise the score. The circularity is confined to the theoretical derivation: Theorem 4.1 promises that the approximate positional encoding at t can represent the future snapshot's positional information without future topology, yet Appendix C's proof requires p^{L+1}=p^1 — an undeclared assumption that the ground-truth encoding of the current/future snapshot equals an old past encoding. This is precisely the kind of future-stationarity the theorem is supposed to establish, so the advertised guarantee is partly equivalent to its own hidden input. I also note two non-circular proof defects: Lemma C.1 states ring-Laplacian eigenvalues as 2−cos(2πk/n) instead of 2−2cos(2πk/n), and the passage from Eq. 20 to Eq. 21 replaces an O(L^2) coefficient sum with 2L−2 without justification. These undermine the theorem as a mathematical result but are correctness errors, not circularity. Overall, the central theoretical prediction is partially circular (score 6), while the experimental findings remain independently supported.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The method is a standard learned model plus hand-chosen hyperparameter families (L, K, t_gap per dataset), two loss weights, and a fixed time encoder; the theoretical claim additionally imports the slowly changing premise and a filter normalization condition, which together already imply the bound it derives. No new physical or formal entities are introduced; the novel mechanism (the learnable spectral filter) is trained end-to-end and is not an unexplained postulate.

free parameters (8)
  • L (PE history length) = 10 to 200 per dataset (Table 22)
    Hand-chosen per dataset; defines the sequence length in Eqs. 1-2 and enters the epsilon bound of Theorem 4.1; the ablation in Table 18 shows performance varies with L.
  • K (number of recent interactions sampled) = 10 to 30 per dataset (Table 22)
    Hand-chosen per dataset; used in Eqs. 3, 5, and 7 for positional and feature aggregation; ablations in Tables 16-17 show moderate sensitivity.
  • t_gap (node feature time window) = 2 to 2000 per dataset (Table 22)
    Hand-chosen per dataset; used in Eq. 4; Tables 14-15 show it matters (best 1000 for Enron, 10 for UN Vote).
  • alpha_pe (weight of PE loss) = 0.5
    Hand-chosen loss weight in Eq. 13; ablation in Table 20 shows a swing of about one point across values.
  • alpha_neg (negative sample weight in PE loss) = 0.3
    Hand-chosen weight in Eq. 12; Table 19 shows robustness across values.
  • alpha, beta (time encoder hyperparameters) = alpha = 10, beta = 10
    Fixed hyperparameters of f_T (Section 2); not tuned per dataset.
  • Spectral filter constraint = sum_i |(Lambda W_filter)_i| < 1
    Imposed in the proof (Appendix C, after Eq. 21) to make the bound small; the paper does not verify that trained filters satisfy it.
  • Model dimensions d_T, d_N, d_E, d_P = 172 (d_P = 72 for Social Evo.)
    Fixed dimensions chosen in Appendix I.2; d_N and d_E are coupled to dataset feature dimensions.
assumptions (5)
  • ad hoc to paper Ground-truth Laplacian positional encodings are approximately constant over adjacent snapshots (p0 approx ... approx p^{t-1}).
    This 'slowly changing' premise of Theorem 4.1 (Section 4) is exactly the stability the theorem concludes for the approximate encodings; the paper provides no evidence the 13 datasets satisfy it.
  • standard math The DFT basis diagonalizes the Laplacian of a ring graph with eigenvalues lambda_k = 2 - cos(2 pi k / n) (Lemma C.1).
    Standard spectral graph theory gives 2 - 2 cos(2 pi k / n) for the unnormalized ring Laplacian; as stated the lemma has a factor-of-two discrepancy, and the proof relies on this diagonalization in Eqs. 17-18.
  • ad hoc to paper The learnable filter and pooling weights satisfy sum_i |(Lambda W_filter)_i| < 1 and sum alpha_j = 1.
    Appendix C, around Eqs. 20-21; these choices are what make epsilon small, and the paper does not verify them for trained filters.
  • domain assumption Aggregating interactions with the same timestamp into snapshots preserves the information needed for link prediction.
    Appendix I.3; standard practice in the temporal graph literature, reasonable for evaluation but changes the granularity of PE updates to per-batch instead of per-snapshot.
  • domain assumption Laplacian eigenvectors of the initial snapshot form an adequate base positional encoding for all later timestamps.
    Section 3.1 initialization and Appendix H.6 robustness test (Random Walk PE on Enron only); generalization to all datasets is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learnable Spatial-Temporal Positional Encoding for Link Prediction." pith.science (2026). https://pith.science/paper/NODXWMTB

@misc{pith2026250608309,
  author       = {Pith},
  title        = {Pith review of: Learnable Spatial-Temporal Positional Encoding for Link Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NODXWMTB}},
  note         = {Machine review of arXiv:2506.08309}
}
read the original abstract

Accurate predictions rely on the expressiveness power of graph deep learning frameworks like graph neural networks and graph transformers, where a positional encoding mechanism has become much more indispensable in recent state-of-the-art works to record the canonical position information. However, the current positional encoding is limited in three aspects: (1) most positional encoding methods use pre-defined, and fixed functions, which are inadequate to adapt to the complex attributed graphs; (2) a few pioneering works proposed the learnable positional encoding but are still limited to the structural information, not considering the real-world time-evolving topological and feature information; (3) most positional encoding methods are equipped with transformers' attention mechanism to fully leverage their capabilities, where the dense or relational attention is often unaffordable on large-scale structured data. Hence, we aim to develop Learnable Spatial-Temporal Positional Encoding in an effective and efficient manner and propose a simple temporal link prediction model named L-STEP. Briefly, for L-STEP, we (1) prove the proposed positional learning scheme can preserve the graph property from the spatial-temporal spectral viewpoint, (2) verify that MLPs can fully exploit the expressiveness and reach transformers' performance on that encoding, (3) change different initial positional encoding inputs to show robustness, (4) analyze the theoretical complexity and obtain less empirical running time than SOTA, and (5) demonstrate its temporal link prediction out-performance on 13 classic datasets and with 10 algorithms in both transductive and inductive settings using 3 different sampling strategies. Also, L-STEP obtains the leading performance in the newest large-scale TGB benchmark. Our code is available at https://github.com/kthrn22/L-STEP.

Figures

Figures reproduced from arXiv: 2506.08309 by the authors.

Figure 1
Figure 1. Overall Framework of L-STEP prediction for now, i.e., only using approximate positional encoding pe t u , we elaborate on how the current approximate positional encoding gets updated and optimized in the next round. After making the current link predictions at time t (as shown in Section 3.2), L-STEP can then “peek” (or “reveal”) the ground-truth links at time t, such that it can update the approximation pe t u into… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 46 canonical work pages

  1. [1]

    Pagerank bandits for link prediction

    Ban, Y., Zou, J., Li, Z., Qi, Y., Fu, D., Kang, J., Tong, H., and He, J. Pagerank bandits for link prediction. In NeurIPS, 2024

    work page 2024

  2. [2]

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

    Chen, J., Gao, K., Li, G., and He, K. Nagphormer: A tokenized graph transformer for node classification in large graphs. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net, 2023. URL https://openreview.net/pdf?id=8KYeilT3Ow

    work page 2023

  3. [3]

    Cong, W., Zhang, S., Kang, J., Yuan, B., Wu, H., Zhou, X., Tong, H., and Mahdavi, M. Do we really need complicated model architectures for temporal networks? In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net, 2023. URL https://openreview.net/pdf?id=ayPPc0SyLv1

    work page 2023

  4. [4]

    N., Ab Hamid, S

    Daud, N. N., Ab Hamid, S. H., Saadoon, M., Sahran, F., and Anuar, N. B. Applications of link prediction in social networks: A review. Journal of Network and Computer Applications, 166: 0 102716, 2020

    work page 2020

  5. [5]

    and Loynd, R

    Diao, C. and Loynd, R. Relational attention: Generalizing transformers for graph-structured tasks. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net, 2023. URL https://openreview.net/pdf?id=cFuMmbWiN6

    work page 2023

  6. [6]

    Dwivedi, V. P. and Bresson, X. A generalization of transformer networks to graphs. CoRR, abs/2012.09699, 2020. URL https://arxiv.org/abs/2012.09699

    arXiv 2012

  7. [7]

    P., Luu, A

    Dwivedi, V. P., Luu, A. T., Laurent, T., Bengio, Y., and Bresson, X. Graph neural networks with learnable structural and positional representations. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022 . OpenReview.net, 2022. URL https://openreview.net/forum?id=wTTjnvGphYj

    work page 2022

  8. [8]

    P., Joshi, C

    Dwivedi, V. P., Joshi, C. K., Luu, A. T., Laurent, T., Bengio, Y., and Bresson, X. Benchmarking graph neural networks. J. Mach. Learn. Res., 24: 0 43:1--43:48, 2023. URL http://jmlr.org/papers/v24/22-0567.html

    work page 2023

Show all 57 references
  1. [9]

    and He, J

    Fu, D. and He, J. DPPIN: A biological repository of dynamic protein-protein interaction network data. In IEEE International Conference on Big Data , 2022

    2022

  2. [10]

    Local motif clustering on time-evolving graphs

    Fu, D., Zhou, D., and He, J. Local motif clustering on time-evolving graphs. In KDD , 2020

    2020

  3. [11]

    I., and He, J

    Fu, D., Fang, L., Maciejewski, R., Torvik, V. I., and He, J. Meta-learned metrics over multi-evolution temporal graphs. In KDD , 2022

    2022

  4. [12]

    Fairness-aware clique-preserving spectral clustering of temporal graphs

    Fu, D., Zhou, D., Maciejewski, R., Croitoru, A., Boyd, M., and He, J. Fairness-aware clique-preserving spectral clustering of temporal graphs. In WWW , 2023

    2023

  5. [13]

    Vcr-graphormer: A mini-batch graph transformer via virtual connections

    Fu, D., Hua, Z., Xie, Y., Fang, J., Zhang, S., Sancak, K., Wu, H., Malevich, A., He, J., and Long, B. Vcr-graphormer: A mini-batch graph transformer via virtual connections. In ICLR , 2024

    2024

  6. [14]

    TGB 2.0: A benchmark for learning on temporal knowledge graphs and heterogeneous graphs

    Gastinger, J., Huang, S., Galkin, M., Loghmani, E., Parviz, A., Poursafaei, F., Danovitch, J., Rossi, E., Koutis, I., Stuckenschmidt, H., Rabbany, R., and Rabusseau, G. TGB 2.0: A benchmark for learning on temporal knowledge graphs and heterogeneous graphs. CoRR, abs/2406.0963...

    2024 doi arXiv

  7. [15]

    D., Coyaud, \'E ., Gon c alves, J., Mojarad, B

    Gupta, G. D., Coyaud, \'E ., Gon c alves, J., Mojarad, B. A., Liu, Y., Wu, Q., Gheiratmand, L., Comartin, D., Tkach, J. M., Cheung, S. W., et al. A dynamic protein interaction landscape of the human centrosome-cilium interface. Cell, 163 0 (6): 0 1484--1499, 2015

    2015

  8. [16]

    Temporal heterogeneous graph generation with privacy, utility, and efficiency

    He, X., Fu, D., Tong, H., Maciejewski, R., and He, J. Temporal heterogeneous graph generation with privacy, utility, and efficiency. In ICLR , 2025

    2025

  9. [17]

    Link prediction approach to collaborative filtering

    Huang, Z., Li, X., and Chen, H. Link prediction approach to collaborative filtering. In Proceedings of the 5th ACM/IEEE-CS joint conference on Digital libraries, pp.\ 141--142, 2005

    2005

  10. [18]

    M., Goel, R., Jain, K., Kobyzev, I., Sethi, A., Forsyth, P., and Poupart, P

    Kazemi, S. M., Goel, R., Jain, K., Kobyzev, I., Sethi, A., Forsyth, P., and Poupart, P. Representation learning for dynamic graphs: A survey. J. Mach. Learn. Res., 21: 0 70:1--70:73, 2020. URL http://jmlr.org/papers/v21/19-447.html

    2020

  11. [19]

    Pure transformers are powerful graph learners

    Kim, J., Nguyen, D., Min, S., Cho, S., Lee, M., Lee, H., and Hong, S. Pure transformers are powerful graph learners. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Advances in Neural Information Processing Systems 35: Annual Conference on Ne...

    2022

  12. [20]

    S., Singh, K., and Biswas, B

    Kumar, A., Singh, S. S., Singh, K., and Biswas, B. Link prediction techniques, applications, and performance: A survey. Physica A: Statistical Mechanics and its Applications, 553: 0 124289, 2020

    2020

  13. [21]

    Predicting dynamic embedding trajectory in temporal interaction networks

    Kumar, S., Zhang, X., and Leskovec, J. Predicting dynamic embedding trajectory in temporal interaction networks. In Teredesai, A., Kumar, V., Li, Y., Rosales, R., Terzi, E., and Karypis, G. (eds.), Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discov...

    2019

  14. [22]

    Crafting papers on machine learning

    Langley, P. Crafting papers on machine learning. In Langley, P. (ed.), Proceedings of the 17th International Conference on Machine Learning (ICML 2000), pp.\ 1207--1216, Stanford, CA, 2000. Morgan Kaufmann

    2000

  15. [23]

    Graphlingo: Domain knowledge exploration by synchronizing knowledge graphs and large language models

    Le, D., Zhao, K., Wang, M., and Wu, Y. Graphlingo: Domain knowledge exploration by synchronizing knowledge graphs and large language models. In 40th IEEE International Conference on Data Engineering, ICDE 2024, Utrecht, The Netherlands, May 13-16, 2024 , pp.\ 5477--5480. IEEE ...

    2024

  16. [24]

    Knowledge graphs can be learned with just intersection features

    Le, D., Zhong, S., Liu, Z., Xu, S., Chaudhary, V., Zhou, K., and Xu, Z. Knowledge graphs can be learned with just intersection features. In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024 . OpenReview.net, 2024 b . URL htt...

    2024

  17. [25]

    Everything evolves in personalized pagerank

    Li, Z., Fu, D., and He, J. Everything evolves in personalized pagerank. In WWW , 2023

    2023

  18. [26]

    Can graph neural networks learn language with extremely weak text supervision? CoRR, 2024

    Li, Z., Zheng, L., Jin, B., Fu, D., Jing, B., Ban, Y., He, J., and Han, J. Can graph neural networks learn language with extremely weak text supervision? CoRR, 2024

    2024

  19. [27]

    Apex\( ^ 2 \): Adaptive and extreme summarization for personalized knowledge graphs

    Li, Z., Fu, D., Ai, M., and He, J. Apex\( ^ 2 \): Adaptive and extreme summarization for personalized knowledge graphs. In KDD , 2025 a

    2025

  20. [28]

    F., Tong, H., and He, J

    Li, Z., Lin, X., Liu, Z., Zou, J., Wu, Z., Zheng, L., Fu, D., Zhu, Y., Hamann, H. F., Tong, H., and He, J. Language in the flow of time: Time-series-paired texts weaved into a unified temporal narrative. CoRR, 2025 b

    2025

  21. [29]

    Backtime: Backdoor attacks on multivariate time series forecasting

    Lin, X., Liu, Z., Fu, D., Qiu, R., and Tong, H. Backtime: Backdoor attacks on multivariate time series forecasting. In NeurIPS, 2024

    2024

  22. [30]

    Class-imbalanced graph learning without class rebalancing

    Liu, Z., Qiu, R., Zeng, Z., Yoo, H., Zhou, D., Xu, Z., Zhu, Y., Weldemariam, K., He, J., and Tong, H. Class-imbalanced graph learning without class rebalancing. In ICML , 2024

    2024

  23. [31]

    and Zhou, T

    Lu, L. and Zhou, T. Link prediction in complex networks: A survey. CoRR, abs/1010.0725, 2010. URL http://arxiv.org/abs/1010.0725

    2010 arXiv

  24. [32]

    Mart \' nez, V., Berzal, F., and Talavera, J. C. C. A survey of link prediction in complex networks. ACM Comput. Surv. , 49 0 (4): 0 69:1--69:33, 2017. doi:10.1145/3012704. URL https://doi.org/10.1145/3012704

    2017 doi

  25. [33]

    Transformer for graphs: An overview from architecture perspective

    Min, E., Chen, R., Bian, Y., Xu, T., Zhao, K., Huang, W., Zhao, P., Huang, J., Ananiadou, S., and Rong, Y. Transformer for graphs: An overview from architecture perspective. CoRR, abs/2202.08455, 2022. URL https://arxiv.org/abs/2202.08455

    2022 arXiv

  26. [34]

    Attending to graph transformers

    M \" u ller, L., Galkin, M., Morris, C., and Ramp \' a sek, L. Attending to graph transformers. CoRR, abs/2302.04181, 2023. doi:10.48550/ARXIV.2302.04181. URL https://doi.org/10.48550/arXiv.2302.04181

    2023 doi arXiv

  27. [35]

    Towards better evaluation for dynamic link prediction

    Poursafaei, F., Huang, S., Pelrine, K., and Rabbany, R. Towards better evaluation for dynamic link prediction. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Advances in Neural Information Processing Systems 35: Annual Conference on Neural I...

    2022

  28. [36]

    P., Luu, A

    Ramp \' a sek, L., Galkin, M., Dwivedi, V. P., Luu, A. T., Wolf, G., and Beaini, D. Recipe for a general, powerful, scalable graph transformer. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Advances in Neural Information Processing Systems ...

    2022

  29. [37]

    Rossi, E., Chamberlain, B., Frasca, F., Eynard, D., Monti, F., and Bronstein, M. M. Temporal graph networks for deep learning on dynamic graphs. CoRR, abs/2006.10637, 2020. URL https://arxiv.org/abs/2006.10637

    2006 arXiv

  30. [38]

    H., Mesquita, D., Kaski, S., and Garg, V

    Souza, A. H., Mesquita, D., Kaski, S., and Garg, V. Provably expressive temporal graph networks. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A. (eds.), Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Pro...

    2022

  31. [39]

    The discrete Fourier transform: theory, algorithms and applications

    Sundararajan, D. The discrete Fourier transform: theory, algorithms and applications. World Scientific, 2001

    2001

  32. [40]

    W., Linding, R., Warde-Farley, D., Liu, Y., Pesquita, C., Faria, D., Bull, S., Pawson, T., Morris, Q., and Wrana, J

    Taylor, I. W., Linding, R., Warde-Farley, D., Liu, Y., Pesquita, C., Faria, D., Bull, S., Pawson, T., Morris, Q., and Wrana, J. L. Dynamic modularity in protein interaction networks predicts breast cancer outcome. Nature biotechnology, 27 0 (2): 0 199--204, 2009

    2009

  33. [41]

    Freedyg: Frequency enhanced continuous-time dynamic graph model for link prediction

    Tian, Y., Qi, Y., and Guo, F. Freedyg: Frequency enhanced continuous-time dynamic graph model for link prediction. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net, 2024. URL https://openreview.net...

    2024

  34. [42]

    F., and He, J

    Tieu, K., Fu, D., Zhu, Y., Hamann, H. F., and He, J. Temporal graph neural tangent kernel with graphon-guaranteed. In NeurIPS, 2024

    2024

  35. [43]

    Invariant link selector for spatial-temporal out-of-distribution problem

    Tieu, K., Fu, D., Wu, J., and He, J. Invariant link selector for spatial-temporal out-of-distribution problem. In The 28th International Conference on Artificial Intelligence and Statistics, 2025

    2025

  36. [44]

    Dyrep: Learning representations over dynamic graphs

    Trivedi, R., Farajtabar, M., Biswal, P., and Zha, H. Dyrep: Learning representations over dynamic graphs. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019 . OpenReview.net, 2019. URL https://openreview.net/forum?id=HyePrhR5KX

    2019

  37. [45]

    N., Kaiser, L., and Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. Attention is all you need. In Guyon, I., von Luxburg, U., Bengio, S., Wallach, H. M., Fergus, R., Vishwanathan, S. V. N., and Garnett, R. (eds.), Advances in Neural Inf...

    2017

  38. [46]

    TCL: transformer-based dynamic graph modelling via contrastive learning

    Wang, L., Chang, X., Li, S., Chu, Y., Li, H., Zhang, W., He, X., Song, L., Zhou, J., and Yang, H. TCL: transformer-based dynamic graph modelling via contrastive learning. CoRR, abs/2105.07944, 2021 a . URL https://arxiv.org/abs/2105.07944

    2021 arXiv

  39. [47]

    Learning graph quantized tokenizers

    Wang, L., Hassani, K., Zhang, S., Fu, D., Yuan, B., Cong, W., Hua, Z., Wu, H., Yao, N., and Long, B. Learning graph quantized tokenizers. In ICLR , 2025

    2025

  40. [48]

    Inductive representation learning in temporal networks via causal anonymous walks

    Wang, Y., Chang, Y., Liu, Y., Leskovec, J., and Li, P. Inductive representation learning in temporal networks via causal anonymous walks. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . OpenReview.net, 2021 b . UR...

    2021

  41. [49]

    Inductive representation learning on temporal graphs

    Xu, D., Ruan, C., K \" o rpeoglu, E., Kumar, S., and Achan, K. Inductive representation learning on temporal graphs. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020 . OpenReview.net, 2020. URL https://openreview....

    2020

  42. [50]

    Language models are graph learners

    Xu, Z., Hassani, K., Zhang, S., Zeng, H., Yasunaga, M., Wang, L., Fu, D., Yao, N., Long, B., and Tong, H. Language models are graph learners. CoRR, 2024

    2024

  43. [51]

    Towards better dynamic graph learning: New architecture and unified library

    Yu, L., Sun, L., Du, B., and Lv, W. Towards better dynamic graph learning: New architecture and unified library. In Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., and Levine, S. (eds.), Advances in Neural Information Processing Systems 36: Annual Conference on Neur...

    2023

  44. [52]

    Drgnn: Deep residual graph neural network with contrastive learning

    Zheng, L., Fu, D., Maciejewski, R., and He, J. Drgnn: Deep residual graph neural network with contrastive learning. Transactions on Machine Learning Research, 2024 a

    2024

  45. [53]

    Pyg-ssl: A graph self-supervised learning toolkit

    Zheng, L., Jing, B., Li, Z., Zeng, Z., Wei, T., Ai, M., He, X., Liu, L., Fu, D., You, J., Tong, H., and He, J. Pyg-ssl: A graph self-supervised learning toolkit. CoRR, 2024 b

    2024

  46. [54]

    Gnns also deserve editing, and they need it more than once

    Zhong, S., Le, D., Liu, Z., Jiang, Z., Ye, A., Zhang, J., Yuan, J., Zhou, K., Xu, Z., Ma, J., Xu, S., Chaudhary, V., and Hu, X. Gnns also deserve editing, and they need it more than once. In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, ...

    2024

  47. [55]

    Mentorgnn: Deriving curriculum for pre-training gnns

    Zhou, D., Zheng, L., Fu, D., Han, J., and He, J. Mentorgnn: Deriving curriculum for pre-training gnns. In CIKM, 2022

    2022

  48. [56]

    GTR: graph-table-rag for cross-table question answering

    Zou, J., Fu, D., Chen, S., He, X., Li, Z., Zhu, Y., Han, J., and He, J. GTR: graph-table-rag for cross-table question answering. CoRR, 2025

    2025

  49. [57]

    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 gl...

Pith tools

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