Pith. sign in

REVIEW 5 major objections 6 minor 43 references

A global graph embedding plus three factored densities jointly predicts where, when, and what the next temporal-graph event is.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 19:11 UTC pith:7SS5M5RP

load-bearing objection Useful joint where/when framing and solid RMSE bake-off vs GNPP, but the headline numbers may be oracle-edge time error and the tables/location eval are unfinished. the 5 major comments →

arxiv 2607.23556 v1 pith:7SS5M5RP submitted 2026-07-26 cs.SI cs.AI

GTIN: A Unified Framework for Joint Event and Time Prediction in Temporal Graphs

Mohammad Ostadmohammadi , Sepehr Kazemi , Hamid R. Rabiee This is my paper
classification cs.SI cs.AI
keywords temporal graphsjoint event-time predictionmarked point processmessage passingglobal graph embeddinghierarchical clusteringDeep Graph Infomax
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

In temporal graphs the usual task is local: guess the next interaction on a given edge or in a small neighborhood. This paper argues that real systems require a global answer—the single next event anywhere in the graph, together with its time and features. It casts the problem as a marked point process whose marks are the edge, the continuous time, and the event attributes, then factors the density as p(edge | history) × p(time | edge, history) × p(features | time, edge, history). History is summarized by treating every edge as a node, running temporal message passing with a harmonic time encoder, and collapsing the resulting edge embeddings into one weighted graph vector. That vector drives a softmax over edges, a Gaussian over time, and a Gaussian over features. On Wikipedia, Reddit, CollegeMsg and synthetic Hawkes/Poisson streams the resulting model reports lower time RMSE than a dozen baselines, with the largest gains on dense, bursty data. A later DGI-plus-balanced-clustering stage partitions large graphs so the same predictor can run in parallel with only modest accuracy loss.

Core claim

The next event anywhere in a temporal graph can be predicted jointly—location, time and features—by a single global edge-level embedding that feeds three factored conditional densities, and this construction yields lower time-prediction error than existing local temporal-graph and point-process baselines on both real and synthetic data.

What carries the argument

Global Temporal Interaction Network (GTIN): each original edge is treated as a node; a harmonic time encoder plus message passing produces edge embeddings that are summed into one graph embedding h_G(t); three heads then output p(c|H), N(μ_t,σ_t²) and N(μ_W,diag σ_W²).

Load-bearing premise

Inter-event times and features are well enough described by diagonal Gaussians whose means and variances are simple functions of the summed global and edge embeddings.

What would settle it

On a held-out stream whose inter-event times are heavy-tailed or multi-modal, replace the Gaussian time head with a heavier-tailed or mixture density; if the reported RMSE advantage over GNPP and TGN disappears or reverses, the joint-prediction claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Next-event forecasts in social, financial or traffic graphs can be produced from a single global state rather than per-edge local models.
  • The same three-factor likelihood can be reused for any mark type once the appropriate output head is substituted.
  • Balanced subgraph clustering plus DGI embeddings lets the identical predictor scale to graphs too large for full-graph message passing.
  • Ablating attention or switching to weaker embeddings measurably degrades both time and location metrics, confirming that the global aggregation step is load-bearing.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the Gaussian time head is the main limitation, swapping it for a neural intensity or mixture density should recover accuracy on sparse or bursty streams without changing the global embedding.
  • The same edge-as-node construction could be applied to continuous-time knowledge-graph completion, turning link prediction into a joint (subject, relation, object, time) marked process.
  • Because the model already emits a full predictive density, calibration and uncertainty-aware downstream decisions (e.g., when to intervene) become immediate next experiments.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The manuscript proposes GTIN, a temporal-graph model that treats edges as nodes of a transformed (line) graph, runs temporal message passing over them, summarizes global history as a weighted sum of edge embeddings, and factors the next-event density as p(c|H) p(t|c,H) p(W|t,c,H) with a softmax location head and diagonal-Gaussian time/feature heads. Training is teacher-forced maximum likelihood on chronological splits; inference selects a (t, W, c) triplet by argmax of the joint density. The paper reports time-RMSE improvements over THP/SAHP/NNPP, several temporal GNNs, and GNPP on Wikipedia, Reddit, CollegeMsg, and three synthetic Hawkes/Poisson graphs (Tables 1–2); Hits@q/MRR for next-edge prediction with no baselines (Table 3); an activity-quartile diagnostic against GNPP (§5.7); and a DGI + balanced k-means partitioning extension that yields 2–3× runtime speedups (Table 4). The problem formulation is reasonable and the factored-density design is clean, but the headline empirical claim rests on an evaluation protocol that is never pinned down, and two of the three claimed prediction targets are weakly or not at all evaluated.

Significance. If the evaluation protocol is clarified and the RMSE gains hold under joint (non-oracle) inference, the paper makes a useful contribution: joint where/when/what prediction on temporal graphs is genuinely underexplored relative to link prediction, and the edge-centric line-graph message passing with a global history embedding is a reasonable and scalable design. The activity-quartile analysis vs GNPP (§5.7) is a commendable, falsifiable diagnostic that ties the claimed mechanism (cross-edge information flow) to an observable signature. The hierarchical DGI + balanced-clustering extension with measured 2–3× speedups is practically relevant. Against this, the evidence base is currently thin: no variance reporting, corrupted table cells, no location baselines, no feature-prediction results, and no code or configuration release, so reproducibility is not established.

major comments (5)
  1. [§5.3 and Tables 1–2] This is the central unresolved issue. §5.3 defines RMSE over t_i − t̂_i but never states whether t̂_i = μ_t conditioned on the true edge c_i (teacher-forced, as in training per §4.3) or on the edge selected by the argmax joint inference rule. The distinction is load-bearing: Table 3 reports Hits@1 of 0.09–0.22, so an oracle-edge protocol would condition time prediction on information unavailable at inference for 78–91% of test events, measuring 'time given true location' rather than the claimed joint where/when task. If baselines were evaluated without equivalent oracle information, Tables 1–2 are not a like-for-like comparison. The manuscript must state the protocol explicitly and, ideally, report RMSE under the joint-inference regime (and its sensitivity to location errors).
  2. [§5.4–5.5, Tables 1–2] All results appear to be single runs: no seeds, error bars, or significance tests are reported. Key margins are small relative to typical run-to-run variance for these benchmarks — Wikipedia 10.6 vs GNPP 11.14 (~5%) and CollegeMsg 27.12 vs 29.75 (~9%) — while the abstract claims the method 'consistently outperforms existing techniques.' Without variance estimates, the consistency claim on two of three real datasets is not supported. Multi-seed means ± std, or at minimum a significance statement, are needed.
  3. [§5.6 and §4.3 (feature head)] The paper claims to jointly predict where, when, and what (§1 contributions; the three-factor likelihood of §4.3), but the experiments only evaluate time. Table 3 reports GTIN's Hits@q/MRR with no baseline whatsoever (GNPP or a simple frequency/recency baseline would be natural), so the 'where' component is unbenchmarked. The 'what' component, p(W|t,c,H) and loss L_W, receives no evaluation at all — no feature RMSE or likelihood is reported on any dataset. Either add these experiments or scope the claims to what is demonstrated.
  4. [Tables 1–2 and §5.1] §4.3 states the Gaussian time head 'can be applied either to normalized timestamps or to normalized inter-event times' without saying which was used per dataset, and no normalization convention or time unit is given. The baseline RMSEs span implausibly different scales (SAHP: 361.07 Wikipedia vs 427.2 Reddit vs GTIN's 0.51 on Reddit), strongly suggesting inconsistent units or normalization across rows. Since RMSE is only meaningful within a fixed convention, Tables 1–2 cannot be audited for fairness until units, normalization, and the per-baseline adaptation protocol (most baselines are link-prediction/embedding models with no native time head) are specified.
  5. [Table 2] The SAHP row reads '11.622.8612.89' and the GNPP row '2.428.212.50' — multiple column values merged without separators. The GNPP row is the strongest-baseline comparison underlying §5.5's claim of '8–70% error reduction,' which is therefore unverifiable as printed. The table must be corrected; given the ambiguity (e.g., is Hawkes-pos SAHP 2.86?), the authors should re-derive the row from their logs rather than guess.
minor comments (6)
  1. [§5, §6.5–6.7, Tables 1 and 4] Broken/truncated cross-references and sentences: §5 opens with 'experiments and results related to.'; §6.5 has 'as illustrated in 2' and 'as shown in 3'; §6.7 has 'Our results (4) show'. These appear to be failed \ref commands. Also Table 1 header 'CollegMsg' (sic), and Table 4's speedup row runs values together ('2.56×3.13×2.08×').
  2. [§4.2] In §4.2 the factorization writes the final event as (t_{NN}, W_{NN}, c_K), mixing subscripts; the history likelihood also contains an extra next-event term log p(t̂,Ŵ,ĉ|H(t)) whose role in training is unclear — during teacher forcing the (K+1)-th event is unobserved, so it is unclear whether this term is actually optimized or is notational only. Please clarify.
  3. [§5.7 / Figure 1] The activity-quartile diagnostic vs GNPP (Fig. 1) is shown only for the Hawkes-positive synthetic data. Since the argument is that cross-edge information drives the gains, showing the same quartile analysis on at least one real dataset (e.g., Reddit, where the largest RMSE gain is claimed) would substantially strengthen §5.7.
  4. [§5.3 and experimental setup] The candidate edge set C_i in §5.3 is never defined — is the softmax over all N edges, or a sampled subset? This affects both the difficulty of Table 3 and the cost claims of §6. Similarly, train/validation/test split ratios, hyperparameters (λ_t, λ_W, d_ω, layers), and hardware for Table 4 are not reported, and no code release is mentioned.
  5. [§6.5–6.7] §6.5 asserts the DGI pipeline 'remains superior to all baseline methods' with only a marginal RMSE increase, but no numerical values are given (Figures 2–4 are unlabeled bar charts in the text). A small table of GTIN vs GTIN+DGI RMSE per dataset would make the scalability/accuracy trade-off claim checkable. Also, the ablation in §6.7 (mean vs attention aggregation) states conclusions about 'nodes in denser regions' without supporting numbers.
  6. [§4.1 and References] Citation style is inconsistent: 'The [1] considers...' (§4.1) uses a numbered citation in an author–year manuscript, and several references list arXiv preprints of published works (e.g., TGAT, ICLR 2020; TGN, ICML 2020 GRL+ workshop; CTDNE, WWW 2018 companion). Please update to venue citations. Also §4.1's second display defines H_n(t,L) with the inequality 't ≤ t_i^(n)' which appears inverted relative to its description as 'the next L interactions.'

Circularity Check

0 steps flagged

No significant circularity: standard factored NLL model trained on chronological history and evaluated on held-out events.

full rationale

The paper’s load-bearing chain is a conventional supervised construction, not a closed derivation. Preliminaries restate the marked point-process likelihood; §4.1–4.2 define the joint next-event objective as the usual factorization p(c|H)p(t|c,H)p(W|t,c,H); §4.3 parameterizes those factors with message-passing edge embeddings, a global sum, softmax, and diagonal Gaussians, then maximizes the negative log-likelihood under teacher forcing. Test metrics (RMSE, Hits@q, MRR) are computed on chronologically held-out events against external baselines (TGAT, TGN, GNPP, etc.). Nothing in that chain equates a fitted constant to a claimed prediction, imports a uniqueness theorem from the same authors, or renames a known empirical pattern as a first-principles result. Self-citations are ordinary background (Hawkes, TGAT, TGN, Xia et al. GNPP) and are not load-bearing for any uniqueness claim. Evaluation-protocol ambiguity (oracle edge vs. model-selected edge for time RMSE) is a correctness/reporting issue, not circularity of the derivation. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 1 invented entities

The central claim rests on standard point-process likelihood identities, several strong modeling choices (Gaussian marks/times, edge homogeneity, global sum pooling), and many fitted neural parameters. No new physical entity is postulated; “GTIN” is an architecture name. Free parameters dominate: network weights, harmonic frequencies, loss weights, and clustering knobs.

free parameters (6)
  • All neural weights (W_Φ, W_c, MLP_t, MLP_W, message-passing AGG/COMB, attention/edge weights w_n)
    Fitted end-to-end to maximize the joint event likelihood; the reported RMSE gains are entirely dependent on this fit.
  • Harmonic time-encoder frequencies ω_1..ω_dω
    Learnable frequencies in ϕ_h(t); control which temporal scales the model can represent.
  • Loss weights λ_t and λ_W
    Trade off classification vs time vs feature terms in L = L_c + λ_t L_t + λ_W L_W; not specified numerically.
  • Gaussian softplus floor ε and time/feature normalization scheme
    Numerical stability and whether absolute time or inter-event time is modeled; changes the meaning of RMSE.
  • DGI encoder/discriminator parameters and balanced k-means k, ε tolerance
    Control subgraph partition quality and the accuracy–runtime tradeoff in §6; fitted or chosen without sensitivity tables.
  • Neighbor activity window Δ for quartile analysis
    Defines Q1–Q4 splits in §5.7; analysis conclusions depend on this hand-chosen window.
axioms (6)
  • standard math Marked temporal point process likelihood: log L = Σ log λ(t_i,k_i) − ∫ λ, equivalently factored densities p(c)p(t|c)p(W|t,c).
    §3–4.2; classical TPP identities used as the training objective.
  • domain assumption Next-graph-event prediction reduces to repeated one-step prediction of H(t,1).
    §4.1 Definition; authors note multi-step needs but train/eval one-step only.
  • ad hoc to paper Conditional time and real-valued features are diagonal-Gaussian given embeddings.
    §4.3; modeling choice without heavy-tail or mixture alternatives or calibration checks.
  • domain assumption Edges are homogeneous and node attributes can be ignored; graph history is a weighted sum of edge embeddings.
    §4.3 pooling; limitations §7.1–7.2 admit this may drop context.
  • domain assumption Neighbor relation on edges = share an endpoint in the original graph; multi-layer MP captures needed global context.
    §4.3 message-passing construction.
  • ad hoc to paper Balanced DGI clusters preserve enough temporal-structural dependence that per-subgraph GTIN approximates full-graph prediction.
    §6; scalability claim depends on this unproven preservation.
invented entities (1)
  • GTIN (Global Temporal Interaction Network) architecture no independent evidence
    purpose: Name for the edge-as-node MP + global sum embedding + factored where/when/what heads.
    Not a new physical object; an engineering assembly of known TPP and GNN pieces. No independent evidence beyond this paper’s tables.

pith-pipeline@v1.2.0-grok45-kimik3 · 19596 in / 4018 out tokens · 85660 ms · 2026-07-30T19:11:16.279440+00:00 · methodology

0 comments
read the original abstract

Temporal graphs are increasingly used to model dynamic systems in diverse domains such as social networks, financial networks, and traffic networks. Predicting both what the next event will be and when it will occur in these systems is crucial for understanding and anticipating complex behaviors, but has not been studied much. To address this gap, we propose a unified mathematical framework capable of capturing varying degrees of complexity across temporal graphs. Our framework is flexible and expressive enough to accommodate a wide range of network structures and temporal dynamics. Building upon this analysis, we introduce our novel approach for jointly predicting the next event and its occurrence time. Empirical evaluations across multiple datasets demonstrate that our method consistently outperforms existing techniques, particularly in scenarios involving irregular event patterns and complex temporal dependencies. These findings highlight the potential of our framework as a robust foundation for future research in temporal event prediction.

Figures

Figures reproduced from arXiv: 2607.23556 by Hamid R. Rabiee, Mohammad Ostadmohammadi, Sepehr kazemi.

Figure 1
Figure 1. Figure 1: Performance (RMSE) of GTIN and GNPP according to recent neighboring-edge activity in the Hawkes-pos dataset. Test events are divided into quartiles based on the number of events observed on neighboring edges during the preceding interval of length Δ [PITH_FULL_IMAGE:figures/full_fig_p016_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of RMSE after adding the pipeline M. Ostadmohammadi et al.: Preprint submitted to Elsevier Page 16 of 15 [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Different clustering methods [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of removing attention M. Ostadmohammadi et al.: Preprint submitted to Elsevier Page 17 of 15 [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 7 linked inside Pith

  1. [3]

    , author Li, Y

    author Gao, S. , author Li, Y. , author Zhang, X. , author Shen, Y. , author Shao, Y. , author Chen, L. , year 2024 . title Simple: Efficient temporal graph neural network training at scale with dynamic data placement . journal Proceedings of the ACM on Management of Data volume 2 , pages 1--25

    2024

  2. [5]

    , year 1971

    author Hawkes, A.G. , year 1971 . title Spectra of some self-exciting and mutually exciting point processes . journal Biometrika volume 58 , pages 83--90

    1971

  3. [7]

    , author Zhang, X

    author Kumar, S. , author Zhang, X. , author Leskovec, J. , year 2019 . title Predicting dynamic embedding trajectory in temporal interaction networks , in: booktitle Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , pp. pages 1269--1278

    2019

  4. [8]

    , author Jiang, X

    author Li, X. , author Jiang, X. , author Zhou, Q. , year 2025 . title Commonsense knowledge enhanced event graph representation learning for script event prediction . journal Machine Learning volume 114 , pages 76

    2025

  5. [9]

    , author Feng, S

    author Li, Z. , author Feng, S. , author Shi, J. , author Zhou, Y. , author Liao, Y. , author Yang, Y. , author Li, Y. , author Yu, N. , author Shao, X. , year 2023 . title Future event prediction based on temporal knowledge graph embedding. journal Computer Systems Science & Engineering volume 44

    2023

  6. [10]

    , author Tu, Z

    author Liu, M. , author Tu, Z. , author Xu, X. , author Wang, Z. , year 2023 . title Who should i engage with at what time? a missing event-aware temporal graph neural network . journal IEEE Transactions on Neural Networks and Learning Systems

    2023

  7. [11]

    , author Eisner, J.M

    author Mei, H. , author Eisner, J.M. , year 2017 . title The neural hawkes process: A neurally self-modulating multivariate point process . journal Advances in neural information processing systems volume 30

    2017

  8. [12]

    , author Lee, J.B

    author Nguyen, G.H. , author Lee, J.B. , author Rossi, R.A. , author Ahmed, N.K. , author Koh, E. , author Kim, S. , year 2018 . title Continuous-time dynamic network embeddings , in: booktitle Companion proceedings of the the web conference 2018 , pp. pages 969--976

    2018

  9. [13]

    , author Aihara, K

    author Omi, T. , author Aihara, K. , et al., year 2019 . title Fully neural network based model for general temporal point processes . journal Advances in neural information processing systems volume 32

    2019

  10. [14]

    , author Opsahl, T

    author Panzarasa, P. , author Opsahl, T. , author Carley, K.M. , year 2009 . title Patterns and dynamics of users' behavior and interaction: Network analysis of an online community . journal Journal of the American Society for Information Science and Technology volume 60 , pages 911--932

    2009

  11. [16]

    , author Wu, Y

    author Sankar, A. , author Wu, Y. , author Gou, L. , author Zhang, W. , author Yang, H. , year 2020 . title Dysat: Deep neural representation learning on dynamic graphs via self-attention networks , in: booktitle Proceedings of the 13th international conference on web search and data mining , pp. pages 519--527

    2020

  12. [17]

    , author Farajtabar, M

    author Trivedi, R. , author Farajtabar, M. , author Biswal, P. , author Zha, H. , year 2019 . title Dyrep: Learning representations over dynamic graphs , in: booktitle International conference on learning representations

    2019

  13. [19]

    , author Li, Y

    author Xia, W. , author Li, Y. , author Li, S. , year 2022 . title Graph neural point process for temporal interaction prediction . journal IEEE Transactions on Knowledge and Data Engineering volume 35 , pages 4867--4879

    2022

  14. [21]

    , author Lipani, A

    author Zhang, Q. , author Lipani, A. , author Kirnap, O. , author Yilmaz, E. , year 2020 . title Self-attentive hawkes process , in: booktitle International conference on machine learning , organization PMLR . pp. pages 11183--11193

    2020

  15. [22]

    , author Wang, X

    author Zheng, T. , author Wang, X. , author Feng, Z. , author Song, J. , author Hao, Y. , author Song, M. , author Wang, X. , author Wang, X. , author Chen, C. , year 2023 . title Temporal aggregation and propagation graph neural networks for dynamic representation . journal IEEE Transactions on Knowledge and Data Engineering volume 35 , pages 10151--10165

    2023

  16. [23]

    , author Kang, Z

    author Zhou, W.T. , author Kang, Z. , author Liu, S. , author Zhang, L. , author Tian, L. , year 2025 . title Fine-grained spatio-temporal event prediction with self-adaptive anchor graph , in: booktitle Proceedings of the 2025 SIAM International Conference on Data Mining (SDM) , organization SIAM . pp. pages 558--567

    2025

  17. [24]

    , author Jiang, H

    author Zuo, S. , author Jiang, H. , author Li, Z. , author Zhao, T. , author Zha, H. , year 2020 . title Transformer hawkes process , in: booktitle International conference on machine learning , organization PMLR . pp. pages 11692--11702

    2020

  18. [25]

    , author Liu, G

    author Zuo, Y. , author Liu, G. , author Lin, H. , author Guo, J. , author Hu, X. , author Wu, J. , year 2018 . title Embedding temporal network via neighborhood formation , in: booktitle Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining , pp. pages 2857--2866

    2018

  19. [26]

    Biometrika , volume=

    Spectra of some self-exciting and mutually exciting point processes , author=. Biometrika , volume=. 1971 , publisher=

    1971

  20. [27]

    Advances in neural information processing systems , volume=

    The neural hawkes process: A neurally self-modulating multivariate point process , author=. Advances in neural information processing systems , volume=

  21. [28]

    International conference on learning representations , year=

    Dyrep: Learning representations over dynamic graphs , author=. International conference on learning representations , year=

  22. [29]

    IEEE Transactions on Neural Networks and Learning Systems , year=

    Who should I engage with at what time? A missing event-aware temporal graph neural network , author=. IEEE Transactions on Neural Networks and Learning Systems , year=

  23. [30]

    arXiv preprint arXiv:2412.12385 , year=

    Enhancing Temporal Link Prediction with HierTKG: A Hierarchical Temporal Knowledge Graph Framework , author=. arXiv preprint arXiv:2412.12385 , year=

    Pith/arXiv arXiv

  24. [31]

    , author=

    Future Event Prediction Based on Temporal Knowledge Graph Embedding. , author=. Computer Systems Science & Engineering , volume=

  25. [32]

    Proceedings of the 2025 SIAM International Conference on Data Mining (SDM) , pages=

    Fine-grained Spatio-temporal Event Prediction with Self-adaptive Anchor Graph , author=. Proceedings of the 2025 SIAM International Conference on Data Mining (SDM) , pages=. 2025 , organization=

    2025

  26. [33]

    Machine Learning , volume=

    Commonsense knowledge enhanced event graph representation learning for script event prediction , author=. Machine Learning , volume=. 2025 , publisher=

    2025

  27. [34]

    arXiv preprint arXiv:2405.18106 , year=

    A unified temporal knowledge graph reasoning model towards interpolation and extrapolation , author=. arXiv preprint arXiv:2405.18106 , year=

    Pith/arXiv arXiv

  28. [35]

    arXiv preprint arXiv:2409.05477 , year=

    Retrofitting temporal graph neural networks with transformer , author=. arXiv preprint arXiv:2409.05477 , year=

    Pith/arXiv arXiv

  29. [36]

    Proceedings of the ACM on Management of Data , volume=

    Simple: Efficient temporal graph neural network training at scale with dynamic data placement , author=. Proceedings of the ACM on Management of Data , volume=. 2024 , publisher=

    2024

  30. [37]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    Temporal aggregation and propagation graph neural networks for dynamic representation , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2023 , publisher=

    2023

  31. [38]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    Graph neural point process for temporal interaction prediction , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2022 , publisher=

    2022

  32. [39]

    Companion proceedings of the the web conference 2018 , pages=

    Continuous-time dynamic network embeddings , author=. Companion proceedings of the the web conference 2018 , pages=

    2018

  33. [40]

    Proceedings of the 13th international conference on web search and data mining , pages=

    Dysat: Deep neural representation learning on dynamic graphs via self-attention networks , author=. Proceedings of the 13th international conference on web search and data mining , pages=

  34. [41]

    Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

    Embedding temporal network via neighborhood formation , author=. Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

  35. [42]

    arXiv preprint arXiv:2003.13432 , year=

    Graph hawkes neural network for forecasting on temporal knowledge graphs , author=. arXiv preprint arXiv:2003.13432 , year=

    Pith/arXiv arXiv 2003

  36. [43]

    arXiv preprint arXiv:2006.10637 , year=

    Temporal graph networks for deep learning on dynamic graphs , author=. arXiv preprint arXiv:2006.10637 , year=

    Pith/arXiv arXiv 2006

  37. [44]

    arXiv preprint arXiv:2002.07962 , year=

    Inductive representation learning on temporal graphs , author=. arXiv preprint arXiv:2002.07962 , year=

    Pith/arXiv arXiv 2002

  38. [45]

    Advances in neural information processing systems , volume=

    Fully neural network based model for general temporal point processes , author=. Advances in neural information processing systems , volume=

  39. [46]

    International conference on machine learning , pages=

    Transformer hawkes process , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  40. [47]

    International conference on machine learning , pages=

    Self-attentive Hawkes process , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  41. [48]

    Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

    Predicting dynamic embedding trajectory in temporal interaction networks , author=. Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

  42. [49]

    Journal of the American Society for Information Science and Technology , volume=

    Patterns and dynamics of users' behavior and interaction: Network analysis of an online community , author=. Journal of the American Society for Information Science and Technology , volume=. 2009 , publisher=

    2009

  43. [50]

    arXiv preprint arXiv:1809.10341 , year=

    Deep graph infomax , author=. arXiv preprint arXiv:1809.10341 , year=

    Pith/arXiv arXiv