Pith. sign in

REVIEW 3 major objections 5 minor 21 references

Back to All-Entity Ranking: Sampler-Dependent Evaluation in Continuous-Time Dynamic Graphs

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Sampled-negative evaluation can flip which CTDG model is ranked best, so model-superiority conclusions are conditional on the sampler and candidate count.

desk verdict A careful, honest paper showing sampled-negative MRR in CTDGs is protocol-conditional; the main empirical flips are real but partly configuration-dependent, and the fixed-catalog recommendation is well scoped. read the letter →

arxiv 2607.27861 v1 pith:5W5PJZCF submitted 2026-07-30 cs.AI

Minwoo Yu , Young-guk Ha This is my paper · ORCID
classification cs.AI
keywords continuous-timedynamicgraphsnegativesamplingall-entityrankingtemporallinkpredictionMRRevaluationpair-historymembershipmodelcomparisoncandidateprotocol
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 MRR scores on continuous-time dynamic graph benchmarks depend on the negative-sampling protocol — the distribution q and the number of candidates K — not just on the model being evaluated. It derives the Bayes-optimal score for sampled ranking as the density ratio log p+(v|x) - log q(v|x), so any non-uniform sampler changes the ranking target, while a uniform finite candidate set sees only a nonlinear partial view of the full catalog rank. Empirically, across six models and four datasets, at least one model pair reverses order between expected Uniform-20 and full-catalog MRR on three datasets, and the CRAFT-R module gain on Wikipedia changes from +0.0037 at K=20 to -0.0266 over the full catalog. The mechanism is pair-history membership: models that use source-destination history directly transmit the sampler's seen/unseen composition into the score. The paper recommends all-entity ranking over a fixed catalog as the primary evidence for architecture comparisons.

What carries the argument

The load-bearing objects are three. First, the density-ratio score ℓ*_q(x,v) = log p+(v|x) - log q(v|x): this is the Bayes-optimal ranking for a sampled candidate set, showing that a non-uniform negative distribution changes what optimal ranking means. Second, the hypergeometric expectation E[RR_K | r, N] (Eq. 7): it maps each event's full-catalog rank r into expected finite-K MRR, isolating the pure effect of K on the metric and removing seed variance. Third, the minimal pair-membership scorer β(M) = b + βM, whose optimal coefficient β*_q = logit(p1) - logit(q1) quantifies how much of sampled MRR can be earned just by knowing whether the pair has been seen before. Together these show how pa

What would settle it

Re-run the six models on the four datasets with the same checkpoints and scorers and compute both expected Uniform-20 and full-catalog MRR; if no model pair reverses relative order on any dataset and the CRAFT-R module gain keeps the same sign for every K and the full catalog, the paper's central empirical claim fails. A cheaper check: on the released code, verify whether any test event's sampled candidate set can contain a positive that is unseen in the training prefix yet still scored against the fixed catalog, which would expose a risk-set mismatch.

Watch

Extended reading notes

Core claim

The central claim is that a sampled-negative MRR is a joint product of the scoring function and the test-time candidate protocol, so model-superiority and ablation conclusions are conditional on q and K. The paper derives the Bayes-optimal score for sampled ranking as log p+(v|x) - log q(v|x), shows that a uniform q over a fixed catalog preserves the population-optimal order while finite-K MRR is only a partial view, and proves the expected finite-K MRR is a hypergeometric quantity determined by the positive's full rank. Pair-history membership — whether the source-destination pair has appeared before — can nearly separate positives from negatives in small random candidate sets. With identic

Load-bearing premise

The whole comparison assumes the right risk set is the fixed catalog of destinations seen in the training prefix; if the deployment setting admits new destinations after the prefix, both sampled MRR and the proposed all-entity MRR are measuring a target that is not the test-time ranking.

Editorial extensions

If this is right

  • Reported model rankings from random-20 or uniform-20 CTDG benchmarks should be read as statements about that candidate configuration, not about the architectures in general.
  • Ablation studies that add a pair-history or recurrence module should report full-catalog MRR, because the module's measured gain can change sign with K and training objective.
  • All-entity ranking over a fixed training catalog removes negative-selection freedom and sampling variance while retaining the original scorer, making architecture comparisons stable and reproducible.
  • Support-restricted samplers such as source-history sampling can make a new positive untestable by construction, since the positive lies outside the sampler support while negatives lie inside it.
  • Expected finite-K MRR computed from full ranks via Eq. (7) is a seed-free metric that separates candidate-draw variance from the K-dependent transformation.

Reading between the lines

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

  • If open-world arrivals matter, neither sampled MRR nor training-prefix all-entity MRR is the deployment target; the paper's recommendation implicitly assumes the destination catalog stops growing at the training prefix, so applying it to Wikipedia/Reddit-style streams requires a time-available or expanding-catalog risk set.
  • The density-ratio view suggests the negative sampler is part of the task definition; one could invert the paper's logic and tune q to define a preferred ranking target, making 'which q' a modeling choice rather than an evaluation nuisance.
  • The representation-intervention result invites a testable extension: if pair-membership direction is removed from CRAFT-R's pre-scorer, retrained models should show reduced Repeated-Unseen amplitude, and all-entity MRR should become a stronger predictor of downstream ranking quality.
  • Since the training objective changes the sign of the module effect (BCE vs BPR on Wikipedia), architecture comparisons should use a unified objective and treat the objective as a confound, not just the evaluation protocol.
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

3 major / 5 minor

Summary. This paper studies how the choice of negative-sampling protocol affects MRR-based evaluation of continuous-time dynamic graph (CTDG) link predictors. It derives the Bayes-optimal ranking under a sampled candidate set (Eqs. 2-3), shows that uniform finite-K evaluation introduces an exact expected-reciprocal-rank transformation of the full-catalog rank (Eqs. 6-7), and analyzes a minimal pair-membership scorer (Eq. 10). The empirical section compares six models on LastFM, MOOC, Reddit, and Wikipedia under a unified BCE training recipe, computing expected Uniform-20 MRR and full-catalog MRR from the same score vectors. The paper reports model-order reversals on three datasets, module-effect sign changes for CRAFT-R versus CRAFT, and a 2x2 membership diagnostic. It concludes that sampled-negative model-superiority and ablation conclusions are conditional on the candidate configuration and recommends all-entity ranking as primary evidence for fixed-catalog benchmarks.

Significance. If the empirical claims hold, this is a valuable contribution to CTDG evaluation methodology. The theoretical core is clean and parameter-free: Eq. (7) gives an exact, closed-form relationship between an event's full-catalog rank and its expected MRR under uniform K-negative sampling, and Eq. (10) provides a simple analytic benchmark for the pair-membership signal. The 2x2 diagnostic and the representation projection are useful tools for separating membership-aligned signals from fine-grained pair compatibility. The recommendation to report all-entity MRR alongside sampled protocols is actionable and directly addresses a real source of protocol-dependent conclusions. The paper also ships reproducible code and a public data-materialization pipeline, which strengthens confidence. The main weakness is that the headline empirical demonstration rests on a non-standard training recipe and three seeds; the observed model-order reversals and module-effect flips are not yet shown to be robust across training objectives and seed realizations. The theoretical and mechanistic parts are solid; the empirical part needs additional support or a narrowing of the claim.

major comments (3)
  1. [§V-A, §V-G, Table IX] All six models are trained with a single non-standard recipe: 3 epochs, BCE, one random negative per positive, and three seeds. This is not the original training configuration for several models (e.g., CRAFT uses BPR in its original study). Table IX shows that the key Wikipedia module-effect sign flip — CRAFT-R minus CRAFT under Expected-20 (+0.0037) versus all-entity (−0.0266) under BCE — becomes +0.0345 and +0.0629 under BPR. Thus the empirical demonstration that sampled versus full-catalog evaluation flips a module conclusion is currently conditional on the BCE training objective. Since the abstract states that the results 'establish' that model-superiority and ablation conclusions are conditional on the candidate configuration, this load-bearing empirical claim needs to be either established under the original training recipes with more seeds, or explicitly scoped to the unified BCE
  2. [§V-E, Table V, Table VI] The central empirical result — 'at least one model pair changes relative order on three of the four datasets' — is inferred from mean MRR values over three seeds. No seed-paired confidence intervals or permutation tests are reported for the model-pair order differences. For example, on LastFM the reversal between CRAFT and DyGFormer is CRAFT .4283±.0112 vs. DyGFormer .4161±.0015 under Expected-20, and .0636±.0065 vs. .0811±.0100 under the full catalog. With n=3, this could be a seed artifact. Table VIII provides seed-paired intervals only for CRAFT-R minus CRAFT, not for the other model-pair reversals. Since the headline claim is empirical, the paper should provide paired evidence on the distribution of the sign of the model-order differences across seeds, or at least a paired bootstrap over seeds. Without this, the 'one pair flips on three datasets' statement is not statistically establ
  3. [§IV-C, §VI] The 'all-entity' risk set is fixed to destinations appearing in the training prefix, and test events whose positive destination is not in that set are excluded (catalog coverage 0.987–1.000). For Wikipedia and Reddit, new destinations continue to arrive after the training prefix, so neither the sampled protocol nor the fixed-catalog protocol directly corresponds to an open-world deployment target. The paper acknowledges this in Section I and Section VI, and the formal claim is scoped to 'an exactly enumerable fixed catalog.' However, the abstract and title omit this qualifier, and the general recommendation to use all-entity ranking as primary evidence could be misread as applying to open-world settings. The discrepancy between the fixed-catalog frame and the actual data-generating process should be stated more prominently, particularly because the paper uses Wikipedia and Reddit as two
minor comments (5)
  1. [Appendix A, Eq. (15)] The displayed formula for Eq. (15) in the appendix has a formatting issue: the numerator is written as `{ N \choose K+1} - {N-r \choose K+1}/{r {N-1 \choose K}}`, which is ambiguous and likely missing parentheses. The intended expression should be `[(N \choose K+1) - (N-r \choose K+1)] / [r (N-1 \choose K)]`.
  2. [Table III] The column header 'History β' is not defined in the table. It should be labeled 'source-history β*' or 'β*_hist' to distinguish it from a coefficient computed from event history in a different sense. Also, the exact equality between analytic and learned β under random sampling is striking and should be stated explicitly as a check of Eq. (10) rather than left implicit.
  3. [Abstract, §V-G] The abstract's 'These results establish ...' is stronger than the evidence supports, given the training-objective sensitivity shown in Table IX. Softening to 'provide evidence' or 'demonstrate under the unified BCE recipe' would better match the content.
  4. [Figure 2] The label 'All-entity decrease' on the right of the horizontal axis is ambiguous. It would be clearer to mark the full-catalog point for Wikipedia explicitly, since the sign change on Wikipedia is one of the main findings.
  5. [§II-B] The reference to Krichene and Rendle and Dallmann et al. is appropriate, but the discussion could briefly note that the CTDG setting adds time-dependent pair membership, which is the novel mechanism. This is implied but would help the reader.

Circularity Check

1 steps flagged · score 1.0 of 10

One minor self-definitional comparison in Table III; central derivation and empirical reversal are self-contained and not circular.

  1. self definitional [Section V-B / Table III / Eq. (10)]
    "Table III tests Eq. (10) under paired-random and source-history conditions on the same events. The analytic β∗ and directly learned β agree on every dataset under random sampling."

    For a logistic regression with a single binary predictor M and an intercept, the maximum-likelihood slope is exactly logit(P(Y=1|M=1)) - logit(P(Y=1|M=0)) = logit(p1) - logit(q1), which is precisely Eq. (10). The analytic β* and the 'directly learned' β are therefore the same quantity computed from the same events and the same p1/q1 statistics. The agreement is an algebraic identity, not an empirical confirmation. The table's framing as a test of Eq. (10) is thus self-definitional: the predicted and learned coefficients share identical inputs by construction. This step is not load-bearing for the paper's central claim about model-order reversals.

full rationale

The paper's load-bearing derivation is not circular. Eq. (3), the Bayes-optimal score under sampling, follows directly from the posterior formula in Eq. (2) under the stated assumptions of known q and equal prior placement of the positive. Eq. (7) is an exact hypergeometric expectation, derived in Appendix A from Eq. (6) with no fitted parameters; it maps a full-catalog rank to the expected finite-K MRR. The central empirical demonstration — that model order and module gain change between expected Uniform-20 and full-catalog evaluation — is explicitly presented as applying Eq. (7) to the per-event full ranks (Section IV-C: 'The primary K-sweep applies Eq. (7) to each event’s full rank to compute the expected MRR under uniform sampling'). The reversal is therefore a disclosed mathematical consequence of the same score vector, not an independently measured quantity being passed off as a prediction. No load-bearing self-citation exists; references to JODIE, DGB, TGB, CRAFT, etc. are contextual or external. The only genuine circularity is the Table III comparison: the analytic coefficient in Eq. (10) is the closed-form logistic-regression MLE for the single binary membership predictor, so its agreement with the learned coefficient is a tautology. This step is minor, does not support the central recommendation, and does not affect the paper's main empirical or theoretical content. Score 1.

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

The central analysis is free of fitted parameters: no constant was tuned to produce the order flips, and the theory is parameter-free from stated assumptions. However, the empirical magnitudes and the specific flips are conditional on hand-chosen protocol values (3-epoch BCE training, history length 320, the K sweep, off-diagonal negative count), and the whole framework presupposes the fixed training-prefix catalog as the risk set and the full-catalog rank as the intended target. No new entities are introduced; the representation intervention and the 2x2 design are analysis tools, not postulated quantities.

free parameters (6)
  • Training epochs = 3 = 3
    Table II. Unified recipe for all six models; far below typical reported training for these architectures. The absolute MRR values and the specific order flips are functions of this choice. Disclosed in Section V-A and partially mitigated by the BPR retraining in Table IX.
  • History length = 320 = 320
    Table II. History truncation used for all models; changes membership prevalence p1/q1 and the 2x2 cell compositions. Chosen by hand; affects the quantitative diagnostics but not the sign of the mechanism.
  • Training negative count (BCE) = 1 per positive
    Table II. The training objective and its sampler remain part of the model specification; the paper shows in Table IX that the objective changes the measured module effect, which is itself one of the paper's supporting results.
  • Off-diagonal cell negatives = 20
    Section IV-A. The 2x2 off-diagonal diagnostics use 20 negatives per condition. The amplitudes (0.888-0.987 vs 0.048-0.508) are large enough that the qualitative conclusion is robust, but the cell values are K-dependent by design.
  • Bootstrap block size = 128
    Table VIII. Time-block paired bootstrap groups test events into blocks of 128; interval endpoints scale mildly with the block choice.
  • Candidate-set sweep K = 1, 5, 20, 50, 100
    Section IV-C. This is the paper's independent variable, not a fitted value; included for completeness of the ledger.
assumptions (6)
  • domain assumption The positive is placed uniformly at random among the K+1 candidate positions in the sampled evaluation set.
    Section III-A. Standard in NCE and sampled-metric analyses; if the placement prior were non-uniform, Eq. (3)'s density-ratio form would pick up an extra position-dependent factor.
  • domain assumption The negative distribution q(v|x) is known exactly and negative draws are independent for Eq. (2); the without-replacement variant uses the joint set sampler Q in Eq. (4).
    Section III-A. The Bayes-optimal score is defined relative to a stated q; in practice the benchmark must report the sampler for the derivation to apply.
  • domain assumption The fixed training-prefix destination catalog R is the correct test-time risk set for 'all-entity' ranking.
    Section IV-C ('The primary risk set R is fixed to all destination identities that occur at least once in the chronological training prefix'). Coverage is 0.987-1.000 on the four datasets, but if the deployment regime is open-world the fixed-catalog target is misspecified. The paper scopes its recommendation to enumerable fixed catalogs (Section VI).
  • domain assumption The intended evaluation target is the full-catalog ranking; sampled MRR is a proxy whose finite-K view is an artifact to be removed.
    Section III-B ('Averaging a nonlinear rank metric across scorers can make not only its absolute value but also relative model order and module gain depend on K'). If researchers only care about sampled-set performance, the critique is moot; the recommendation presupposes the full-catalog rank is the quantity of interest.
  • standard math Standard combinatorial identities: Vandermonde convolution for Eq. (7) and the tie-aware hypergeometric sum in Eq. (16); the saturated one-parameter logistic MLE equals the logit-difference of empirical rates.
    Appendix A and Section III-C. Verified by inspection; the derivations are correct.
  • domain assumption Pair membership M(u,v,t) is defined with respect to history Ht up to time t, and source and destination roles are distinct.
    Eq. (9) and Section I. The analysis excludes homogeneous graphs with interchangeable roles; the paper states this scoping explicitly.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Back to All-Entity Ranking: Sampler-Dependent Evaluation in Continuous-Time Dynamic Graphs." pith.science (2026). https://pith.science/paper/5W5PJZCF

@misc{pith2026260727861,
  author       = {Pith},
  title        = {Pith review of: Back to All-Entity Ranking: Sampler-Dependent Evaluation in Continuous-Time Dynamic Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5W5PJZCF}},
  note         = {Machine review of arXiv:2607.27861}
}
read the original abstract

Next-destination prediction in continuous-time dynamic graphs (CTDGs) commonly ranks an observed interaction against sampled negative destinations. The resulting score is conditional on both the negative distribution and the number of candidates chosen by the researcher. We show that a non-uniform negative distribution changes the Bayes-optimal ranking, while even a finite candidate set drawn uniformly can destabilize model rankings and measured module effects. Time-varying source-destination history membership and model operations that use this information directly transmit the sampler's influence to the evaluation score. We examine this mechanism using a factorial evaluation of repeated and new positives against seen and unseen negatives, a minimal scorer based solely on pair-history membership, and controlled representation interventions. Across six models on LastFM, MOOC, Reddit, and Wikipedia, at least one model pair changes relative order between the expected Uniform-20 metric and the full catalog on three of the four datasets. The measured effect of the same module also changes in magnitude and direction with the candidate-set size and training objective. These results establish that model-superiority and ablation conclusions from sampled-negative benchmarks are conditional on the stated candidate configuration. All-entity ranking evaluates every destination in a fixed catalog, eliminating negative-selection freedom and sampling variation while retaining the original CTDG scorer. We therefore recommend all-entity ranking as the primary evidence for architecture comparisons on CTDG benchmarks with an enumerable, fixed destination catalog.

Figures

Figures reproduced from arXiv: 2607.27861 by the authors.

Figure 1
Figure 1. Pair-membership 2 × 2 MRR for CRAFT-R. Rows identify the positive pair and columns identify the negative pair by historical membership. All four cells are directional diagnostics using 20 negatives. TABLE IV MEMBERSHIP-CONTROLLED ALL-SEEN AND ALL-UNSEEN MRR. CRAFT CRAFT-R Dataset All-seen All-unseen All-seen All-unseen LastFM .1939 .0193 .2041 .0145 MOOC .6253 .5780 .6247 .5648 Reddit .8715 .0967 .8549 .1063 Wikiped… view at source ↗
Figure 2
Figure 2. CRAFT-R minus CRAFT MRR as a function of the expected number of negative candidates under uniform sampling. The dashed line marks zero [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. All-entity scoring time and peak GPU memory versus fixed-catalog size. We evaluate 128 test events from each CRAFT checkpoint in candidate [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 2 linked inside Pith

  1. [1]

    Temporal graph networks for deep learning on dynamic graphs,

    E. Rossi, B. Chamberlain, F. Frasca, D. Eynard, F. Monti, and M. Bron- stein, “Temporal graph networks for deep learning on dynamic graphs,” arXiv preprint arXiv:2006.10637, 2020

    arXiv 2006

  2. [2]

    Towards better evaluation for dynamic link prediction,

    F. Poursafaei, S. Huang, K. Pelrine, and R. Rabbany, “Towards better evaluation for dynamic link prediction,” inAdvances in Neural Infor- mation Processing Systems, vol. 35, 2022

    2022

  3. [3]

    Temporal graph benchmark for machine learning on temporal graphs,

    S. Huang, F. Poursafaei, J. Danovitch, M. Fey, W. Hu, E. Rossi, J. Leskovec, M. Bronstein, G. Rabusseau, and R. Rabbany, “Temporal graph benchmark for machine learning on temporal graphs,” inAdvances in Neural Information Processing Systems, vol. 36, 2023

    2023

  4. [4]

    New perspectives on the evaluation of link prediction algorithms for dynamic graphs,

    R. Romero, T. De Bie, and J. Lijffijt, “New perspectives on the evaluation of link prediction algorithms for dynamic graphs,”arXiv preprint arXiv:2311.18486, 2023

    arXiv 2023

  5. [5]

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

    J. Gastinger, S. Huang, M. Galkin, E. Loghmani, A. Parviz, F. Pour- safaei, J. Danovitch, E. Rossi, I. Koutis, H. Stuckenschmidt, R. Rabbany, and G. Rabusseau, “TGB 2.0: A benchmark for learning on temporal knowledge graphs and heterogeneous graphs,” inAdvances in Neural Information Processing Systems, vol. 37, 2024

    2024

  6. [6]

    On sampled metrics for item recom- mendation,

    W. Krichene and S. Rendle, “On sampled metrics for item recom- mendation,” inProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2020, pp. 1748– 1757

    2020

  7. [7]

    A case study on sampling strate- gies for evaluating neural sequential item recommendation models,

    A. Dallmann, D. Zoller, and A. Hotho, “A case study on sampling strate- gies for evaluating neural sequential item recommendation models,” in Proceedings of the 15th ACM Conference on Recommender Systems, 2021, pp. 505–514

    2021

  8. [8]

    Towards better dynamic graph learning: New architecture and unified library,

    L. Yu, L. Sun, B. Du, and W. Lv, “Towards better dynamic graph learning: New architecture and unified library,” inAdvances in Neural Information Processing Systems, vol. 36, 2023

    2023

Show all 21 references
  1. [9]

    Efficient neural com- mon neighbor for temporal graph link prediction,

    X. Zhang, Y . Wang, X. Wang, and M. Zhang, “Efficient neural com- mon neighbor for temporal graph link prediction,” inarXiv preprint arXiv:2406.07926, 2024

    2024

  2. [10]

    Future link prediction without memory or aggregation,

    L. Yi, R. Lei, F. Mo, Y . Zheng, Z. Wei, and Y . Ye, “Future link prediction without memory or aggregation,” inAdvances in Neural Information Processing Systems, vol. 38, 2025

    2025

  3. [11]

    TAMI: Taming heterogeneity in temporal interactions for temporal graph link prediction,

    Z. Yu, J. Wu, Z. Wu, S. Zhong, W. Su, C.-H. Lee, and W. Zhuo, “TAMI: Taming heterogeneity in temporal interactions for temporal graph link prediction,” inAdvances in Neural Information Processing Systems, vol. 38, 2025

    2025

  4. [12]

    Predicting dynamic embedding trajectory in temporal interaction networks,

    S. Kumar, X. Zhang, and J. Leskovec, “Predicting dynamic embedding trajectory in temporal interaction networks,” inProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2019, pp. 1269–1278

    2019

  5. [13]

    Repeat-aware neighbor sampling for dynamic graph learning,

    T. Zou, Y . Mao, J. Ye, and B. Du, “Repeat-aware neighbor sampling for dynamic graph learning,” inProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024

    2024

  6. [14]

    Improving temporal link prediction via temporal walk matrix projection,

    X. Lu, L. Sun, T. Zhu, and W. Lv, “Improving temporal link prediction via temporal walk matrix projection,” inAdvances in Neural Information Processing Systems, vol. 37, 2024

    2024

  7. [15]

    Towards ideal temporal graph neural networks: Evaluations and conclusions after 10,000 gpu hours,

    Y . Yang, H. Zhou, R. Kannan, and V . Prasanna, “Towards ideal temporal graph neural networks: Evaluations and conclusions after 10,000 gpu hours,”Proceedings of the VLDB Endowment, vol. 18, no. 4, pp. 956– 969, 2024

    2024

  8. [16]

    Ranking on dynamic graphs: An effective and robust band-pass disentangled approach,

    Y . Li, Y . Xu, X. Lin, W. Zhang, and Y . Zhang, “Ranking on dynamic graphs: An effective and robust band-pass disentangled approach,” in Proceedings of the ACM Web Conference, 2025

    2025

  9. [17]

    Dyrep: Learning representations over dynamic graphs,

    R. Trivedi, M. Farajtabar, P. Biswal, and H. Zha, “Dyrep: Learning representations over dynamic graphs,” inInternational Conference on Learning Representations, 2019. [Online]. Available: https: //openreview.net/forum?id=HyePrhR5KX

    2019

  10. [18]

    Do we really need complicated model architectures for temporal networks?

    W. Cong, S. Zhang, J. Kang, B. Yuan, H. Wu, X. Zhou, H. Tong, and M. Mahdavi, “Do we really need complicated model architectures for temporal networks?” inInternational Conference on Learning Repre- sentations, 2023

    2023

  11. [19]

    Neighborhood-aware scalable temporal network representation learning,

    Y . Luo and P. Li, “Neighborhood-aware scalable temporal network representation learning,” inLearning on Graphs Conference, 2022, pp. 1–18

    2022

  12. [20]

    Noise-contrastive estimation: A new estimation principle for unnormalized statistical models,

    M. Gutmann and A. Hyv ¨arinen, “Noise-contrastive estimation: A new estimation principle for unnormalized statistical models,” inProceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, 2010, pp. 297–304. 11

    2010

  13. [21]

    Strictly proper scoring rules, prediction, and estimation,

    T. Gneiting and A. E. Raftery, “Strictly proper scoring rules, prediction, and estimation,”Journal of the American Statistical Association, vol. 102, no. 477, pp. 359–378, 2007. APPENDIX Additional Reproducibility Information. Every sampled-Kand full-catalog result is derived ...

    2007

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.