EdgeReMIND: A Scalable, Top-Ranked Memorization Baseline for Temporal Multi-Relational Link Prediction
cs.LG
Quick verdict
Linear per-relation memorization model claims SOTA MRR on 6/8 TGB 2.0 datasets; thorough ablations, but largest-graph wins face no running competitors.
Summary
The paper introduces EdgeReMIND, a CPU-only, embedding-free memorization model for temporal multi-relational link prediction on TGB 2.0. It scores candidate destinations with a per-relation 15-dimensional linear layer over six count/recency features plus nine decay features whose half-lives are placed geometrically ({m/2, m, 2m}) around each scope's median inter-recurrence gap from the training split. The entire learned budget is 15|R| parameters (at most 17,880). Trained with Adam on softmax cross-entropy over K=20 relation-aware negatives and selected by smoothed validation-best-epoch, it reports the highest test MRR on six of eight TGB 2.0 datasets (e.g., thgl-myket 0.906, thgl-github 0.784, tkgl-wikidata 0.640) and is the only relation-aware method running end-to-end on all datasets, including the three largest where every embedding method OOMs or exceeds the seven-day budget. Ablations attribute the largest gains to per-relation learning (+0.34 on thgl-github), with the multi-timescale bank mattering mainly on year-scale TKGs; on thgl-myket and tkgl-icews the untrained recurrence baseline supplies nearly all performance. Main limitations: second place on thgl-software (0.501 vs STHN 0.731) and tkgl-polecat (0.171 vs TLogic 0.228, below RecB), no generalization to novel patterns, hand-set gamma and n, and headline wins on the three largest datasets where no learned competitor has any leaderboard entry.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- EdgeReMIND sets the highest reported test MRR on six of eight TGB 2.0 datasets and is the only relation-aware temporal link-prediction method that runs end-to-end on all eight, CPU-only.
- Per-relation learning of a 15-dimensional weight vector is the primary driver of gains; the data-calibrated multi-timescale bank adds value mainly on year-scale, gap-concentrated TKGs (tkgl-smallpedia +0.151 bank ablation).
- On the three largest TGB 2.0 datasets every existing embedding-based method runs out of memory or exceeds the seven-day budget, so a scalable memorization baseline addresses a genuine production-scale gap.
- A single hand-set configuration with no per-dataset tuning suffices across all datasets, with validation-best-epoch selection and zero-initialized bank columns preserving the untrained recurrence-plus-popularity baseline as a safety floor.
Strengths
Unusual transparency about where memorization, not learning, does the work
Figure 5 shows the untrained baseline supplies 0.9019 of 0.9053 on thgl-myket and 0.2692 of 0.2699 on tkgl-icews; Table 13 shows the fully relation-blind model is within 0.005 on the ret=1.00 datasets. The paper scopes its own claims against the obvious 'it is just EdgeBank' objection rather than hiding it.
Strong leakage control and streaming-conformance argument
The offline index enforces t'<t by prefix binary search (§3.6), and an independent streaming implementation is verified bit-identical on every feature column including all nine bank columns (§B.4.1) — a rigor rarely applied to memorization baselines, where test-history leakage is the classic failure mode.
Broad, mostly five-seed ablation grid
Concrete scalability evidence with full resource accounting
High reproducibility
Weaknesses
Three of the six SOTA wins are wins by default
Substantial losses where real competition exists
Training-evaluation mismatch on 1-vs-all TKGs is patched, not fixed
Internal inconsistency on seed determinism
Central decomposition rests on a single seed
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is EdgeReMIND?
EdgeReMIND is a CPU-only baseline for temporal link prediction on the TGB 2.0 benchmark. Instead of learning node embeddings, it scores candidate destinations with a small linear model — 15 learned weights per relation type — over memorization features such as how often a link was seen before and how recently, evaluated at multiple timescales calibrated from the data's own recurrence gaps. Its largest model has fewer than 18,000 parameters, yet the paper reports the best published test MRR on six of the benchmark's eight datasets.
How is EdgeReMIND different from EdgeBank?
EdgeBank predicts a link only if the same source–destination pair was seen before, optionally within a recent window, using fixed rules. EdgeReMIND extends this pairwise recurrence with relation-conditioned counts and recency features, adds a multi-timescale decay bank whose half-lives are calibrated from the dataset's median recurrence gaps, and learns a separate 15-dimensional weight vector per relation by gradient descent. In the paper's Table 3 it beats both EdgeBank variants on all eight datasets (e.g., 0.784 vs 0.413 on thgl-github).
What MRR does EdgeReMIND achieve on TGB 2.0 datasets?
Five-seed test MRR from the paper: thgl-myket 0.906, thgl-github 0.784, thgl-forum 0.731, thgl-software 0.501, tkgl-wikidata 0.640, tkgl-smallpedia 0.614, tkgl-icews 0.269, tkgl-polecat 0.171. It ranks first on six of the eight datasets; it is second on thgl-software (behind STHN at 0.731) and tkgl-polecat (behind TLogic at 0.228). Note that MRR values are not comparable across datasets because TGB 2.0 uses different negative-sampling regimes per dataset.
Does EdgeReMIND need a GPU?
No. The paper reports all results on CPU only (64 cores, 128 GB RAM, no GPU). Because the model is a single 15-weights-per-relation linear layer over precomputed memorization features, memory grows with the index of past edges rather than with learned representations. Peak memory ranged from 13.6 GB to 73.4 GB per run, and wall-clock from about 2 minutes to 13 hours per seed depending on the dataset.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–6 are available only as extracted text with companion tables; visual content (e.g., Figure 6 gap histograms, Figure 4 scatter layout) could not be independently verified.
- Competitor numbers in Table 3 are taken verbatim from the paper's copy of the official TGB 2.0 leaderboard and were not independently re-checked against the leaderboard itself.
- Several Appendix D factorial cells are marked selector-unstable (*) by the authors; this review treats those values as non-estimates, consistent with the paper's own reading.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.