Pith. sign in

REVIEW 4 major objections 5 minor 26 references

Base3: a simple interpolation-based ensemble method for robust dynamic link prediction

T0 review · 4 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 A training-free ensemble of three simple signals rivals deep temporal graph models on benchmark link prediction.

desk verdict Simple training-free ensemble that looks competitive with deep models on TGB, but the headline comparison rests on test-set-tuned hyperparameters and a missing weights table; needs a validation-based re-run before I'd trust the numbers. read the letter →

arxiv 2506.12764 v2 pith:UONOSCYH submitted 2025-06-15 cs.LG

Kondrup Emma This is my paper · ORCID
classification cs.LG
keywords dynamiclinkpredictiontemporalgraphlearningtraining-freebaselineensembleinterpolationedgerecurrencenodepopularityco-occurrencememoryBenchmark
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 show that complex neural architectures are not necessary for competitive dynamic link prediction on the Temporal Graph Benchmark. It builds a fully training-free, deterministic ensemble called Base3 that linearly combines three non-learnable signals: edge recurrence (EdgeBank), global node popularity (PopTrack), and a newly proposed temporal co-occurrence memory (t-CoMem). The author argues that this simple interpolation reaches state-of-the-art-level Mean Reciprocal Rank, even ranking first on tgbl-coin and close to the top on tgbl-wiki-v2 and tgbl-flight, while requiring no training, backpropagation, or learned parameters. If correct, the central claim matters because it shows that carefully chosen hand-crafted temporal signals can match or beat deep models on several realistic benchmarks, and that the field may be overvaluing architectural complexity.

What carries the argument

The load-bearing mechanism is the interpolation of three complementary non-learnable scorers. EdgeBank stores observed edges and assigns a score of 1 to any edge seen before. PopTrack assigns a binary score based on whether the destination node ranks in the top-$K$ most popular nodes by recent interaction frequency. The new t-CoMem module maintains, for each source node, a time-windowed deque of recently observed destinations; it scores a candidate pair by summing, over all recent neighbors $n_i$ of $u$, the decayed popularity $d \cdot p_i$ with $d = \exp(-(t-t_i)/t_w)$, then adds a co-occurrence term $f = \lambda \cdot c/(1+c)$, where $c$ is the observed co-occurrence count, and finally squashes the sum through $1/(1+1/(\sum d\cdot p_i + f))$. Base3 combines the three scores with heuristic weights chosen by a confidence scheme (multi_conf) that boosts EdgeBank when the edge is in memory and boosts PopTrack when the destination is popular. This machinery carries the argument because it converts two known baselines and one new memory module into a cheap, deterministic predictor whose only tuning is a few scalar hyperparameters.

What would settle it

Re-run Base3 on the full TGB suite with the multi_conf weights shown (the paper's Table ?? in Appendix A is missing) and with per-dataset hyperparameter tuning; if the default configuration loses to a tuned configuration by a large margin on tgbl-review-v2, tgbl-comment, or tgbl-coin, then the reported competitive results do not reflect a generally robust configuration.

Watch

Extended reading notes

Core claim

On the paper's own terms, Base3 is an interpolation model with scores $\alpha \cdot s_{\text{EB}} + \beta \cdot s_{\text{PT}} + \delta \cdot s_{\text{CM}}$, where $s_{\text{EB}}$ marks whether an edge appeared before, $s_{\text{PT}}$ marks whether the destination is among the top-$K$ popular nodes, and $s_{\text{CM}}$ is the new t-CoMem score that combines the source node's recently observed neighbors with decayed PopTrack popularity of those neighbors plus a co-occurrence count. The paper reports test MRR values of 0.743 on tgbl-wiki-v2, 0.108 on tgbl-review-v2, 0.773 on tgbl-coin, 0.450 on tgbl-comment, and 0.794 on tgbl-flight, with t-CoMem alone reaching 0.840 on tgbl-flight. These numbers put Base3 above TGN, GraphMixer, and TNCN on several datasets, and above DyGFormer on tgbl-coin. The claimed discovery is that a deterministic, interpretable combination of recurrence, popularity, and co-occurrence can rival modern deep temporal graph networks, and that the robustness is especially visible under inductive and historical negative sampling, where existing models degrade sharply.

Load-bearing premise

The default hyperparameters (memory span 0.1, co-occurrence weight 1.0, and $K=1000$), along with the weights of the multi_conf interpolation scheme, were fixed using ablation experiments on tgbl-wiki-v2, and the paper assumes these same choices remain near-optimal on the other four datasets.

Editorial extensions

If this is right

  • If the reported numbers hold, training-free methods become the natural first baseline on the Temporal Graph Benchmark, since they are deterministic, fast, and require no hyperparameter learning.
  • On low-to-moderate surprise datasets such as tgbl-wiki-v2 and tgbl-coin, deep models' advantage over simple recurrence and popularity signals appears small or even negative, so those leaderboards may be conflating memorization with generalization.
  • The strong robustness of t-CoMem and Base3 under inductive and historical negative sampling suggests that soft popularity propagation is a more reliable inductive signal than EdgeBank's hard memorization or PopTrack's binary top-$K$ cutoff.
  • The modular design means each component can be reused as a cheap, interpretable building block inside larger learned systems, for example as a prior, an initialization, or a regularizer.

Reading between the lines

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

  • The result is an existence proof that hand-designed temporal signals can saturate much of what deep models extract on these benchmarks; a natural extension is to test the same three-signal combination on other temporal graph suites beyond TGB.
  • The paper leaves implicit that benchmark leaderboards dominated by deep models may partly measure how well networks memorize recurring edges and track popularity, rather than how well they model complex temporal dependencies.
  • A testable extension is to replace the hand-set multi_conf weights with a small logistic regression or a per-dataset learned scalar, which could raise performance further while still keeping most of the interpretability.
  • The paper argues for efficiency but reports no wall-clock runtime; a direct head-to-head timing comparison against TGN-style models would make the deployment claim concrete.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Base3, a training-free ensemble for dynamic link prediction that linearly combines three non-learnable signals: EdgeBank-style edge recurrence, PopTrack-style node popularity, and a new t-CoMem module that tracks temporal co-occurrence patterns with recency decay. The authors evaluate Base3 and t-CoMem on the Temporal Graph Benchmark (TGB) datasets and report MRR scores competitive with deep learning baselines such as DyGFormer, TNCN, TGN, and GraphMixer, including the highest test MRR on tgbl-coin and strong performance on tgbl-wiki-v2 and tgbl-flight. They also report AUROC under random, inductive, and historical negative sampling. The central claim is that a simple, interpretable, training-free model can rival state-of-the-art deep temporal graph models, especially under more challenging negative sampling strategies.

Significance. If the reported results are reliable, the paper would be a valuable contribution to the growing literature on lightweight baselines for temporal graph learning. The proposed method is deterministic, requires no training, and the authors provide code, which are strengths for reproducibility. The evaluation across five TGB datasets and three negative sampling strategies is broad. However, the significance is substantially undercut by two methodological concerns: the default hyperparameters appear to be selected using test-set performance rather than validation performance, and the multi_conf interpolation weights are not fully specified (Appendix A contains a missing table reference). These issues affect the fairness of the comparison with deep baselines and the reproducibility of the method. If the authors can re-establish the results under validation-based model selection and provide complete weight definitions, the contribution would be solid.

major comments (4)
  1. [§4.3, Table 3] The selection of default hyperparameters appears to maximize test MRR rather than validation MRR. For K=1000 and co-occurrence weight 1.0, memory span 1.0 gives validation MRR 0.736 and test MRR 0.727, while span 0.1 gives validation MRR 0.727 and test MRR 0.743. The text states that span 0.1 is fixed "as that is optimal under optimal choices for other hyperparameters," but a validation-based choice would favor span 1.0. Since the deep baselines follow the standard TGB protocol of selecting hyperparameters on validation, this discrepancy gives Base3 an unfair advantage and undermines the claim that it is competitive with state-of-the-art models on the reported test sets.
  2. [§4.4, Table 4] The interpolation strategy multi_conf is chosen based on test-set results. On tgbl-wiki-v2, EB_conf achieves test MRR 0.752 while multi_conf achieves 0.743; the authors nevertheless select multi_conf because it performs better on tgbl-review-v2 (test MRR 0.108 vs 0.084). This means the default configuration of Base3 is selected using test information from two datasets. To fairly compare against deep baselines, the interpolation strategy should be selected using validation MRR or a separate development set, and the reported test numbers should be re-generated under that selection rule.
  3. [Appendix A] The multi_conf interpolation weights are not fully specified. The prose states that when both EdgeBank and PopTrack signals are reliable, t-CoMem receives 20%, and when only one signal is positive, EdgeBank receives 45% or PopTrack receives 70%, but the complete weight vectors (α, β, δ) for all four cases are not given, and the text refers to a "Table ??" that is missing. Without these weights, the method cannot be reproduced exactly, and there is no way to verify whether the weights were chosen by inspecting test results. The authors should provide the complete weight table and clarify how the weights were determined.
  4. [§3.1] The t-CoMem time window tw is a free hyperparameter set to 1,000,000 by default, but it is not included in the ablation study in Table 3. Since t-CoMem is a core component of Base3, the reported results are conditional on an unstated and apparently untuned choice. A sensitivity analysis for tw should be provided, or at minimum the default should be justified.
minor comments (5)
  1. [Table 1 caption] The caption states that first, second, and third best performances are highlighted, but the table as typeset does not show any highlighting; this should be fixed so the reader can verify the rankings.
  2. [Equation (4)] Equation (4) contains a stray "= 1" in the numerator; the formula for the Resource Allocation Index should be cleaned up.
  3. [§4.4] The text uses both "tgbl-review-v2" and "tgbl-review" in Table 4 and the surrounding discussion; the dataset name should be used consistently.
  4. [Introduction] The code link is given as "available here" without an actual URL; the URL should be included in the manuscript.
  5. [Throughout] The term "co-occurence" is misspelled; it should be "co-occurrence" in the text and in the hyperparameter description.

Circularity Check

2 steps flagged · score 6.0 of 10

Reported test MRRs are selected on the test set via hyperparameter and interpolation-model choice, so the headline competitiveness is partially a fitted result.

  1. fitted input called prediction [Section 4.3 (Ablation Studies), Table 3]
    "As such, we find the optimal memory span to be either 0.1 or 1.0, fixing it to 0.1 as that is optimal under optimal choices for other hyperparameters. ... As such, we select the optimal combination of hyperparameters and set Base3's defaults to them – specifically, we set a memory span of 0.1, a co-occurence weight of 1.0 and a K-value of 1000."

    In Table 3, with K=1000 and co-occurrence weight 1.0, the validation MRR at memory span 0.1 is 0.727 while at span 1.0 it is 0.736; only the test MRR favors span 0.1 (0.743 vs 0.727). The paper's stated 'optimal' default is therefore the test-set optimum, not the validation optimum. Consequently the test MRRs reported in Table 1 for this default configuration are selected maxima rather than out-of-sample predictions, so the headline claim 'Base3 achieves a high test MRR (0.743)' reduces, for this dataset, to a test-set fit.

  2. fitted input called prediction [Section 4.4 (Comparing Interpolation Models), Table 4]
    "This trend, however, is not reproduced in other datasets, as we empirically observed, and report for tgbl-review. Indeed, on tgbl-review, both uniform and EB_conf perform a few points lower than multi_conf. Considering tgbl-review is harder than tgbl-wiki-v2 (it has a higher surprise), these insights are quite important, and determined us setting Base3's default strategy to multi_conf to produce a more robust model."

    The choice between interpolation models is made using test MRR on the same datasets where final performance is later reported. In Table 4, EB_conf has a higher test MRR on tgbl-wiki-v2 (0.752 vs 0.743 for multi_conf), but multi_conf is selected because of its test MRR on tgbl-review-v2 (0.108 vs 0.084). Thus the reported Table 1 test results for the default Base3 configuration are the product of model selection on the test sets, not independent predictions of those test sets.

full rationale

The paper does not rest on self-citation: EdgeBank [16] and PopTrack [4] are external prior work, and no uniqueness theorem is imported. The derivation of the three component scores (EdgeBank recurrence, PopTrack popularity, t-CoMem co-occurrence) is self-contained, and the components are not defined in terms of the final MRR. However, the central claim that Base3 is competitive with state-of-the-art deep models is weakened by circular evaluation: the default hyperparameters (memory span 0.1, co-occurrence weight 1.0, K=1000) and the default interpolation scheme (multi_conf) are selected using test-set MRR. Table 3 shows the validation optimum at K=1000 is span 1.0 (val 0.736), while the chosen span 0.1 has val 0.727 but test 0.743, so the stated 'optimal' choice is the test optimum. Table 4 similarly shows EB_conf beating multi_conf on tgbl-wiki-v2 test (0.752 vs 0.743) yet multi_conf is chosen using tgbl-review-v2 test scores. Consequently the reported test numbers are selected maxima rather than out-of-sample predictions, fitting the 'fitted input called prediction' pattern. The missing 'Table ??' for the multi_conf weights and the 'empirically motivated' weight choices make it impossible to rule out further test-set fitting. Because the component heuristics are honest and the issue is confined to evaluation protocol rather than the derivation of the scores, a score of 6 is appropriate.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a small set of hand-selected hyperparameters and interpolation weights, plus the assumption that the evaluation protocol and the cited baselines are sound. No new physical or conceptual entities are postulated; t-CoMem is a software module, not a theoretical entity.

free parameters (6)
  • top-K popularity threshold = 1000
    Set to 1000 after ablation on tgbl-wiki-v2; the paper notes original PopTrack used K=100, but K=1000 gives about 50% higher MRR. Location: Section 4.3.
  • memory span (fraction of history) = 0.1
    Chosen from {0.01, 0.1, 1.0} via ablation on tgbl-wiki-v2; selection appears to consider test MRR. Location: Section 4.3.
  • co-occurrence weight lambda = 1.0
    Chosen from {0.25, 0.5, 0.75, 1.0}; higher values yield higher MRR on tgbl-wiki-v2. Location: Section 4.3.
  • multi_conf interpolation weights = partially unspecified (missing Table ??)
    Hand-set confidence-based weights, 'partially empirically motivated', with four cases for combinations of EB and PT scores. The table is missing in Appendix A. Location: Appendix A.
  • EB_conf interpolation weights = w_conf=[0.5,0.2,0.3], w_not=[0.2,0.3,0.5]
    Hand-set based on 'empirically-motivated preference for t-CoMem'. Location: Appendix A.
  • t-CoMem time window tw = 1,000,000 (default)
    Fixed default, not ablated; used in recency decay exp(-(t-t_i)/tw). Location: Section 3.1.
assumptions (3)
  • domain assumption TGB benchmark evaluation protocol (splits, random negative sampling, MRR) is a valid measure of dynamic link prediction performance
    The paper relies on TGB [8] for all main results; no re-implementation of the protocol is described. Location: Section 4.1.
  • domain assumption EdgeBank and PopTrack are correctly characterized and their reported limitations are accurate
    The method builds directly on the definitions and inductive biases of these cited baselines; the paper does not re-derive them. Location: Sections 2 and 3.
  • domain assumption The three signals (recurrence, popularity, co-occurrence) are complementary across datasets
    The interpolation only helps if the signals are not redundant; the paper presents this as a design hypothesis. Location: Section 3.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Base3: a simple interpolation-based ensemble method for robust dynamic link prediction." pith.science (2026). https://pith.science/paper/UONOSCYH

@misc{pith2026250612764,
  author       = {Pith},
  title        = {Pith review of: Base3: a simple interpolation-based ensemble method for robust dynamic link prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UONOSCYH}},
  note         = {Machine review of arXiv:2506.12764}
}
read the original abstract

Dynamic link prediction remains a central challenge in temporal graph learning, particularly in designing models that are both effective and practical for real-world deployment. Existing approaches often rely on complex neural architectures, which are computationally intensive and difficult to interpret. In this work, we build on the strong recurrence-based foundation of the EdgeBank baseline, by supplementing it with inductive capabilities. We do so by leveraging the predictive power of non-learnable signals from two complementary perspectives: historical edge recurrence, as captured by EdgeBank, and global node popularity, as introduced in the PopTrack model. We propose t-CoMem, a lightweight memory module that tracks temporal co-occurrence patterns and neighborhood activity. Building on this, we introduce Base3, an interpolation-based model that fuses EdgeBank, PopTrack, and t-CoMem into a unified scoring framework. This combination effectively bridges local and global temporal dynamics -- repetition, popularity, and context -- without relying on training. Evaluated on the Temporal Graph Benchmark, Base3 achieves performance competitive with state-of-the-art deep models, even outperforming them on some datasets. Importantly, it considerably improves on existing baselines' performance under more realistic and challenging negative sampling strategies -- offering a simple yet robust alternative for temporal graph learning.

Figures

Figures reproduced from arXiv: 2506.12764 by the authors.

Figure 1
Figure 1. Test MRR by dataset for non-learnable dynamic [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 19 canonical work pages

  1. [1]

    Albert-László Barabási and Réka Albert. 1999. Emergence of scaling in random networks. Science 286, 5439 (1999), 509–512

    1999

  2. [2]

    Weilin Cong, Si Zhang, Jian Kang, Baichuan Yuan, Hao Wu, Xin Zhou, Hang- hang Tong, and Mehrdad Mahdavi. 2023. Do we really need complicated model architectures for temporal networks?. In International Conference on Learning Representations (ICLR)

    work page 2023

  3. [3]

    Filip Cornell, Oleg Smirnov, Gabriela Zarzar Gandler, and Lele Cao. 2025. On the Power of Heuristics in Temporal Graphs. arXiv preprint

    work page 2025

  4. [4]

    Michal Daniluk and Jacek Dabrowski. 2024. Temporal graph models fail to capture global temporal dynamics. arXiv preprint, ICLR 2024 (withdrawn sub- mission). openreview:9kLDrE5rsW Available at https://openreview.net/forum? id=9kLDrE5rsW

    work page 2024

  5. [5]

    Manuel Dileo and Matteo Zignani. 2024. Link prediction heuristics for temporal graph benchmark. In Proceedings of the European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN 2024) . https: //www.esann.org/sites/default/files/proceedings/2024/ES2024-141.pdf

    work page 2024

  6. [6]

    David Easley and Jon Kleinberg. 2010. Networks, Crowds, and Markets: Reasoning About a Highly Connected World . Cambridge University Press

    work page 2010

  7. [7]

    Hamilton, Rex Ying, and Jure Leskovec

    William L. Hamilton, Rex Ying, and Jure Leskovec. 2017. Inductive Represen- tation Learning on Large Graphs. In Advances in Neural Information Processing Systems (NeurIPS), Vol. 30. https://papers.nips.cc/paper_files/paper/2017/hash/ 5dd9db5e033da9c6fb5ba83c7a7ebea9-Abstract.html

    work page 2017

  8. [8]

    Shenyang Huang, Farimah Poursafaei, Jacob Danovitch, Matthias Fey, Weihua Hu, Emanuele Rossi, Jure Leskovec, Michael Bronstein, Guillaume Rabusseau, and Reihaneh Rabbany. 2023. Temporal Graph Benchmark for Machine Learning on Temporal Graphs. Advances in Neural Information Processing Systems (2023)

    work page 2023

Show all 26 references
  1. [9]

    Shenyang Huang, Farimah Poursafaei, Reihaneh Rabbany, Guillaume Rabusseau, and Emanuele Rossi. 2024. UTG: Towards a Unified View of Snapshot and Event Based Models for Temporal Graphs. arXiv preprint. https://arxiv.org/abs/2407. 12269 arXiv:2407.12269 [cs.LG]

    2024 arXiv

  2. [10]

    Seyed Mehran Kazemi, Rishab Goel, Shikhar Jain, Ivan Kobyzev, Luke Sethi, Peter Forsyth, and Pascal Poupart. 2020. Representation learning for dynamic graphs: A survey. Journal of Machine Learning Research 21, 70 (2020), 1–73

    2020

  3. [11]

    Thomas N Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In International Conference on Learning Repre- sentations (ICLR). https://arxiv.org/abs/1609.02907

    2017 arXiv

  4. [12]

    David Liben-Nowell and Jon Kleinberg. 2007. The link-prediction problem for social networks. In Journal of the American society for information science and technology, Vol. 58. Wiley Online Library, 1019–1031

    2007

  5. [13]

    Ryan N Lichtenwalter, Jake T Lussier, and Nitesh V Chawla. 2010. New per- spectives and methods in link prediction. Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining (2010), 243–252

    2010

  6. [14]

    Shahrad Mohammadzadeh. 2024. Temporal Collaborative Filtering: Enhancing EdgeBank with Inductive Capabilities. (2024). Manuscript, not published

    2024

  7. [15]

    Amber et al. Pareja. 2020. EvolveGCN: Evolving Graph Convolutional Networks for Dynamic Graphs. In Proceedings of the AAAI Conference on Artificial Intelli- gence, Vol. 34. 5363–5370

    2020

  8. [16]

    Farimah Poursafaei, Shenyang Huang, Kellin Pelrine, and Reihaneh Rabbany

  9. [17]

    Emanuele Rossi, Ben Chambers, Rex Ying, Michael Bronstein, and Bruno Ribeiro

  10. [18]

    Jo Skarding, Bogdan Gabrys, and Katarzyna Musial. 2021. Foundations and mod- elling of dynamic graphs. In Proceedings of the 2021 International Joint Conference on Neural Networks (IJCNN) . IEEE, 1–8. Base3: a simple interpolation-based ensemble method for robust dynamic link ...

    2021

  11. [19]

    Rakshit Trivedi, Mehrdad Farajtabar, Parnam Biswal, and Hongyuan Zha. 2019. DyRep: Learning Representations over Dynamic Graphs. In International Con- ference on Learning Representations (ICLR) . https://openreview.net/forum?id= HylMyhR5tm

    2019

  12. [20]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2018. Graph Attention Networks. In International Con- ference on Learning Representations (ICLR)

    2018

  13. [21]

    Da Xu, Chuanwei Ruan, Evren Korpeoglu, Sushant Kumar, and Kannan Achan

  14. [22]

    Le Yu, Leilei Sun, Bowen Du, and Weifeng Lv. 2023. Towards Better Dynamic Graph Learning: New Architecture and Unified Library. In Advances in Neural Information Processing Systems, Vol. 36. 67686–67700

    2023

  15. [23]

    Xiaohui Zhang, Yanbo Wang, Xiyuan Wang, and Muhan Zhang. 2024. Ef- ficient Neural Common Neighbor for Temporal Graph Link Prediction. arXiv:2406.07926 [cs.LG] A INTERPOLATION MODELS Here, we provide a more detailed overview of the EB_conf and multi_conf interpolation models, s...

    2024

  16. [24]

    In International Conference on Learning Representations (ICLR)

    Inductive Representation Learning on Temporal Graphs. In International Conference on Learning Representations (ICLR)

  17. [2020]

    arXiv preprint arXiv:2006.10637 (2020)

    Temporal Graph Networks for Deep Learning on Dynamic Graphs. arXiv preprint arXiv:2006.10637 (2020)

    2020 arXiv

  18. [2022]

    arXiv preprint arXiv:2207.10128 (2022)

    Towards Better Evaluation for Dynamic Link Prediction. arXiv preprint arXiv:2207.10128 (2022). https://arxiv.org/abs/2207.10128

    2022 arXiv

Pith tools

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