Pith. sign in

REVIEW 3 major objections 5 minor 63 references

GRATE: Temporal Extensions for Inductive KG Foundation Models via Gated Rotary Attention

T0 review · 3 major / 5 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Temporal reasoning can be added to inductive knowledge-graph foundation models with zero new parameters by rotating edge messages by relative time gaps and gating them by query relevance.

desk verdict Solid parameter-free temporal add-on for inductive KG models, with honest uneven results and a real DRec confound that the authors partly document themselves. read the letter →

arxiv 2607.10197 v1 pith:XS2Y6PRR submitted 2026-07-11 cs.AI

Jiaxin Pan , Osama Mohammed , Daniel Hernández , Steffen Staab This is my paper · ORCID
classification cs.AI
keywords temporalknowledgegraphsinductivetransfergraphfoundationmodelsrotarypositionembeddingquery-conditionedgatinglinkpredictionzero-shotforecastingNBFNet
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

Knowledge-graph foundation models already transfer to graphs with unseen entities and relations by learning structure rather than fixed vocabularies. Temporal knowledge graphs break that transfer: prior temporal models bind parameters to specific entities, relations, or timestamps, so they cannot move to graphs with disjoint vocabularies. This paper claims that a parameter-free entity-side message function, GRATE, is enough to restore temporal judgment while keeping the base model’s structural transferability. It rotates each supporting fact’s message by its time gap to the query and multiplies by a query-conditioned gate that keeps only temporally and semantically aligned signals. A single jointly pretrained checkpoint then improves over the static base model on new inductive transfer suites with fully disjoint entities, relations, and timestamps, and on held-out forecasting sets, in most settings.

What carries the argument

GRATE (Gated Rotary Attention for Temporal Encoding): a parameter-free entity-side message function that rotates each base message by the relative time gap Δt = τ − t_uv via fixed-frequency RoPE, then multiplies by a sigmoid gate measuring alignment between the rotated message and the query state, so only temporally relevant signals enter aggregation.

What would settle it

Train one GRATE checkpoint on daily political-event graphs, evaluate zero-shot on a yearly encyclopaedic graph (and the reverse) under fully disjoint entity/relation/timestamp splits; if relative-time rotation plus gating no longer improves over the static base model once granularities and recurrence patterns diverge, the transfer claim fails.

Watch

Extended reading notes

Core claim

The paper establishes that relative-time rotary encoding plus query-conditioned sigmoid gating can inject temporal similarity into NBFNet-style foundation models without adding any learnable parameters or tying representations to absolute timestamps. Because both operations depend only on relative gaps and existing hidden states, a single checkpoint pretrained on source temporal graphs transfers zero-shot to target graphs whose entities, relations, and timestamps are entirely unseen, covering both interpolation and extrapolation.

Load-bearing premise

The claim rests on the idea that fixed-frequency rotations of raw integer time gaps, with no dataset-level scaling and no explicit handling of different calendar granularities, already produce a transferable notion of temporal relevance across daily and yearly graphs and across past and future prediction.

Editorial extensions

If this is right

  • Inductive KG foundation models can be lifted to temporal link prediction without retraining per entity, relation, or timestamp vocabulary.
  • Cross-dataset temporal transfer becomes measurable: the constructed GDELTIndT and WIKIIndT suites provide V/R/T-disjoint interpolation and extrapolation protocols that prior TKG benchmarks lack.
  • A single jointly pretrained checkpoint can serve multiple held-out forecasting targets without fine-tuning, recovering a large fraction of supervised forecasting performance on some datasets.
  • Gains concentrate where neighborhoods are dense and literal recurrence is concentrated at the preceding timestamp; sparse or diffuse-recurrence graphs leave the gate with less useful signal.
  • Because GRATE adds zero parameters, timestamp-disjoint inductive transfer cannot fail from vocabulary mismatch the way embedding-table methods do.

Reading between the lines

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

  • If relative-gap rotation is the main carrier of temporal signal, mixed-granularity pretraining (daily plus yearly sources remapped to disjoint integer ranges) should further stabilise transfer without new parameters.
  • The same message-function swap could be tried on other NBFNet-style inductive models beyond the two bases tested, testing whether richer relation graphs always amplify the gate’s lift.
  • DRec (fraction of test triples whose exact match sits at the immediately preceding timestamp) appears to predict when GRATE helps; collecting DRec on new targets could serve as a cheap pre-deployment filter for when to enable the module.
  • Parameter-free temporal gating may also help continuous or irregular event streams if absolute clock values are first mapped to ordered integer offsets, an extension the paper leaves open.
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. The paper proposes GRATE, a parameter-free entity-side message function that extends NBFNet-style inductive KG foundation models (ULTRA, TRIX) to temporal knowledge graphs. It encodes relative time gaps Δt = τ − t_uv by rotating base messages with fixed-frequency RoPE and applies a query-conditioned sigmoid gate to reweight rotated messages, adding no learnable parameters. To evaluate V/R/T-disjoint transfer, the authors construct GDELTIndT and WIKIIndT (interpolation and extrapolation, varying p_tri) and also report transductive ICEWS results and zero-shot transfer to ICEWS18 and YAGO. A single jointly pretrained checkpoint on ICEWS14+ICEWS05-15 improves over the corresponding static base models in most settings, with component ablations (Table 3) and a DRec analysis (Fig. 4) offered as supporting evidence.

Significance. Extending vocabulary-agnostic KG foundation models to time without reintroducing entity/relation/timestamp embeddings is a clear and timely problem. The design is attractive: zero added parameters, drop-in replacement of the entity message function, and explicit inductive transfer benchmarks with disjoint entities, relations, and timestamps fill a genuine evaluation gap left by standard TKG splits. Strengths include the parameter-free construction (Eqs. 4–12, Algorithm 1), the public construction protocol for GDELTIndT/WIKIIndT, the component ablation in Table 3, and unusually honest reporting of flat or negative cases (WIKI low-p_tri, ICEWS18 TRIX). If the gains are shown to reflect transferable relative-time reasoning rather than mainly recency-weighted recurrence, the work would be a solid contribution to graph foundation models for temporal data.

major comments (3)
  1. [§7.2–7.3, Fig. 4] §7.2–7.3 and Fig. 4: the reported lift of GRATE over static TRIX is strongly monotone in DRec (literal (s,r,o) match at the immediately preceding timestamp). On high-DRec cells (WIKIIndT-p100, YAGO) the mechanism can act as a recency-biased recurrence amplifier; on low/diffuse DRec (WIKI p25–p50, ICEWS18) gains vanish or reverse. This is load-bearing for the claim of transferable relative-time encoding. Please add controls that separate multi-frequency relative-time geometry from simple recency: e.g., (i) a parameter-free gate that uses only a scalar decay of |Δt| without RoPE rotation, (ii) a shuffle of timestamps within the inference message graph that preserves recurrence counts but destroys ordered Δt structure, and/or (iii) stratified metrics on non-recurrent vs recurrent test facts. Without such controls, the central interpretation remains under-supported relative to a simpler rece
  2. [§5.1, Eqs. (4)–(9), Appendix B.1] §5.1 Eqs. (4)–(9) and Appendix B.1: relative gaps are fed as raw integer Δt with fixed RoPE base 10000 and no dataset-level normalization or granularity handling. Pretraining uses only daily ICEWS sources, while WIKI/YAGO are yearly and GDELT is 15-min in the source (then daily-like in the suite). The Limitations section correctly notes weaker effect under coarse/irregular time, but the main claims still present GRATE as a transferable temporal encoding across these calendars. Either (a) provide evidence that the same fixed basis yields consistent temporal similarity across daily vs yearly integer scales (e.g., controlled synthetic Δt probes, or mixed-granularity pretraining), or (b) narrow the claim to same-granularity transfer and treat cross-granularity results as exploratory.
  3. [§6.2–6.3, Tables 7–8] §6.2–6.3 and Tables 7–8: the inductive suites are a valuable contribution, but the paper’s strongest narrative (“foundation-model temporal transfer”) rests on a single checkpoint pretrained only on ICEWS14+ICEWS05-15. Both sources are daily political-event graphs with similar relation structure. For a journal-level foundation-model claim, the transfer story needs either a broader pretraining mix (including at least one non-ICEWS / non-daily source) or a clearer framing that the current checkpoint is a same-domain inductive temporal adapter rather than a general TKG foundation model. As written, success on GDELT is more persuasive than success on WIKI/YAGO for the foundation-model framing.
minor comments (5)
  1. [Figure 1, Eq. (4)] Figure 1 panel (a) writes Δt_i = t_i − τ in the illustration while Eq. (4) defines Δt_uv = τ − t_uv. The sign convention should be consistent throughout; RoPE is not sign-symmetric in interpretation even if both directions are usable.
  2. [Table 2, §7.3] Table 2 omits MRR for ICL/GenTKG (reasonable if they decode rather than rank) but then compares Hits@1/10 without stating candidate-set size differences clearly enough in the main text; a short note on ranking-all-entities vs pre-filtered decoding would help readers.
  3. [§5.2, Table 3] In §5.2 the gate is called “query-conditioned,” while Table 3 also reports a “relation-conditioned gate.” Define the relation-conditioned variant explicitly (what replaces Q_q) so the ablation row is reproducible from the text alone.
  4. [§6.3, Appendix C.2] Appendix C.2 step 6: the mix of seen vs disjoint edges at p_tri < 1 is important for interpreting “fully inductive” stress. A one-sentence reminder in the main §6.3 that only p_tri=1 is fully R/T-disjoint would reduce ambiguity when reading Figure 2.
  5. [Throughout] Minor typos/notation: “Y AGO” spacing is inconsistent; “GDELTINDT/WIKIINDT” vs “GDELTIndT/WIKIIndT” casing differs between abstract and body; “extramode” missing space in Appendix C.2.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical ML method + transfer evaluation; claims are measured performance deltas on held-out/constructed benchmarks, not derivations that reduce to fitted inputs.

full rationale

GRATE is defined as a parameter-free entity-side message (RoPE rotation of DistMult messages by raw integer Δt plus a query-conditioned sigmoid gate; Eqs. 4–12, Alg. 1) that is independent of any evaluation metric. The central claim is an empirical performance improvement of one jointly pretrained checkpoint over static ULTRA/TRIX bases on V/R/T-disjoint inductive suites (GDELTIndT/WIKIIndT, constructed by an explicit extension of INGRAM) and held-out forecasting sets. Pretraining sources (ICEWS14/05-15) are standard public splits; metrics (time-aware MRR/Hits) are computed on disjoint test facts never seen in training or hyper-parameter selection. Ablations (Table 3) and DRec analyses (Fig. 4) are post-hoc diagnostics, not inputs that force the reported numbers by construction. There are no uniqueness theorems, self-citation load-bearing premises, fitted parameters renamed as predictions, or self-definitional loops. The paper is self-contained against external benchmarks; any debate about whether gains are driven by recency bias rather than transferable temporal encoding is a correctness/interpretation question, not circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

GRATE is deliberately parameter-free; load-bearing content is design choices and domain assumptions about relative time, not fitted constants. Free parameters are standard architecture/training knobs inherited from the base models plus the fixed RoPE base. Invented entities are the method and the two benchmark suites.

free parameters (3)
  • RoPE base b = 10000
    Fixed at 10000 following standard RoPE; not tuned. Still a hand-chosen scale that interacts with raw integer Δt across daily vs yearly data.
  • hidden dimension d and layer count L = d=32, L=6
    Inherited from ULTRA/TRIX configs (d=32, L=6); not free parameters of GRATE itself but required for the reported checkpoints.
  • timestamp remapping ranges for multi-TKG pretraining
    Disjoint integer ranges assigned per source so local indices do not collide; preserves relative gaps but is an author-chosen encoding step (Section 5.3).
assumptions (4)
  • ad hoc to paper Relative integer time gaps Δt = τ − t_uv with fixed multi-frequency rotation suffice to encode temporal relevance without absolute timestamp embeddings or dataset-specific normalization.
    Core design of Section 5.1 and Appendix B.1; Limitations explicitly notes weakness on coarse/irregular granularities.
  • domain assumption Independent per-edge sigmoid gating (not neighborhood softmax) is compatible with NBFNet additive aggregation and more stable under inductive neighborhood-size shift.
    Stated in Section 5.2 as matching NBFNet aggregation structure.
  • domain assumption Structural relation-graph representations from ULTRA/TRIX remain valid when entity-side messages are temporally rotated and gated.
    Integration claim in Section 5 and Figure 1b; GRATE does not modify RelGNN.
  • domain assumption V/R/T-disjoint transfer with controlled p_tri and inter/extra splits is a valid stress test of temporal foundation-model transfer.
    Problem definition Section 3 and construction Section 6.3 / Appendix C.
invented entities (2)
  • GRATE message function (rotary relative-time + query-conditioned gate)
    purpose: Inject temporal displacement and query relevance into entity-side messages without learnable temporal parameters.
    Defined by Equations (4)–(12) and Algorithm 1; evaluated only inside this paper’s experiments.
  • GDELTIndT and WIKIIndT benchmark suites
    purpose: Provide fully inductive temporal transfer evaluation with disjoint entities, relations, and timestamps under interpolation and extrapolation.
    Constructed by the authors via extended INGRAM procedure; primary evidence for the transfer claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GRATE: Temporal Extensions for Inductive KG Foundation Models via Gated Rotary Attention." pith.science (2026). https://pith.science/paper/XS2Y6PRR

@misc{pith2026260710197,
  author       = {Pith},
  title        = {Pith review of: GRATE: Temporal Extensions for Inductive KG Foundation Models via Gated Rotary Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XS2Y6PRR}},
  note         = {Machine review of arXiv:2607.10197}
}
read the original abstract

Knowledge graph foundation models such as Ultra and Trix achieve strong inductive transfer by learning relation-graph representations that generalise to unseen entities and relations. Extending this transferability to temporal knowledge graphs (TKGs) remains challenging: existing temporal models tie their parameters to dataset-specific entities, relations, or timestamps and are not designed to transfer to TKGs with disjoint vocabularies. We propose GRATE (Gated Rotary Attention for Temporal Encoding), an entity-side message function that adds no learnable parameters and encodes time through relative time differences by rotating each edge message according to its time gap to the query and applying a query-conditioned gate to select temporally relevant signals. GRATE integrates into NBFNet-style KG foundation models while preserving structural transferability. Existing TKG benchmarks evaluate within shared train/test vocabularies and cannot directly test cross-dataset temporal transfer; we therefore construct GDELTIndT and WIKIIndT, inductive transfer benchmark suites with disjoint entities, relations, and timestamps spanning both interpolation and extrapolation. Across these benchmarks and held-out forecasting datasets, a single jointly pretrained GRATE checkpoint improves over the static base model in most settings.

Figures

Figures reproduced from arXiv: 2607.10197 by the authors.

Figure 1
Figure 1. GRATE mechanism and integration. (a) Overview of the four processing steps for an example query (Barack Obama, engage in negotiation, ?, Dec 2014). (1) Observed supporting facts from the message graph. (2) Semantic base messages muv = ϕ(hu, ruv) computed by the KG score function. (3) Relative-time rotation: each base message is rotated by R(∆tuv), where ∆tuv = τ − tuv, encoding temporal displacement at multiple freq… view at source ↗
Figure 2
Figure 2. Zero-shot inductive transfer MRR on GDELTINDT and WIKIINDT as a function of ptri. Hits@1/Hits@10 variants are in [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Hits@1 (top) and Hits@10 (bottom) counterparts of [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: GRATE gain (∆ MRR over static TRIX) against DRec across all 16 GDELTINDT/WIKIINDT variants and the two held￾out targets (ICEWS18, YAGO). Each point is labelled by its ptri level. B.2. Base Model Integration GRATE integrates into any NBFNet-style base model as a replace…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 1 linked inside Pith

  1. [1]

    2024 , eprint=

    A Survey on Temporal Knowledge Graph: Representation Learning and Applications , author=. 2024 , eprint=

    2024

  2. [2]

    Neurocomputing , volume=

    Roformer: Enhanced transformer with rotary position embedding , author=. Neurocomputing , volume=. 2024 , publisher=

    2024

  3. [3]

    IJCAI , year=

    History Repeats Itself: A Baseline for Temporal Knowledge Graph Forecasting , author=. IJCAI , year=

  4. [4]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Towards Global-Topology Relation Graph for Inductive Knowledge Graph Completion , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  5. [5]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    GraphOracle: Efficient Fully-Inductive Knowledge Graph Reasoning via Relation-Dependency Graphs , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  6. [6]

    NeurIPS 2023 Workshop: New Frontiers in Graph Learning , year=

    A Multi-Task Perspective for Link Prediction with New Relation Types and Nodes , author=. NeurIPS 2023 Workshop: New Frontiers in Graph Learning , year=

    2023

  7. [7]

    Advances in Neural Information Processing Systems , year=

    TGB 2.0: A Benchmark for Learning on Temporal Knowledge Graphs and Heterogeneous Graphs , author=. Advances in Neural Information Processing Systems , year=

  8. [8]

    Advances in neural information processing systems , volume=

    Attention is all you need , author=. Advances in neural information processing systems , volume=

Show all 63 references
  1. [9]

    Temporal Knowledge Graphs: Integration, Querying, and Analytics , year =

    Jing Ao and Jon Doyle and Christopher Healey and Ranga Raju Vatsavai , advisor =. Temporal Knowledge Graphs: Integration, Querying, and Analytics , year =

  2. [10]

    Question Answering Over Temporal Knowledge Graphs

    Saxena, Apoorv and Chakrabarti, Soumen and Talukdar, Partha. Question Answering Over Temporal Knowledge Graphs. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (V...

    2021 doi

  3. [11]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  4. [12]

    Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval , pages=

    Temporal knowledge graph reasoning based on evolutional representation learning , author=. Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval , pages=

  5. [13]

    Publications Manual , year = "1983", publisher =

    1983

  6. [14]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    1981 doi

  7. [15]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  8. [16]

    Dan Gusfield , title =. 1997

    1997

  9. [17]

    International Conference on Learning Representations , year=

    Tensor Decompositions for Temporal Knowledge Base Completion , author=. International Conference on Learning Representations , year=

  10. [18]

    Advances in neural information processing systems , volume=

    Translating embeddings for modeling multi-relational data , author=. Advances in neural information processing systems , volume=

  11. [19]

    Proceedings of the International Conference on Learning Representations (ICLR) 2015 , year=

    Embedding Entities and Relations for Learning and Inference in Knowledge Bases , author=. Proceedings of the International Conference on Learning Representations (ICLR) 2015 , year=

    2015

  12. [20]

    Proceedings of the 29th ACM International Conference on Information & Knowledge Management , pages=

    Rotate3d: Representing relations as rotations in three-dimensional space for knowledge graph embedding , author=. Proceedings of the 29th ACM International Conference on Information & Knowledge Management , pages=

  13. [21]

    THE WEB CONFERENCE 2025 , year =

    RiemannGFM: Learning a Graph Foundation Model from Structural Geometry , author=. THE WEB CONFERENCE 2025 , year =

    2025

  14. [22]

    Advances in Neural Information Processing Systems , volume=

    Gft: Graph foundation model with transferable tree vocabulary , author=. Advances in Neural Information Processing Systems , volume=

  15. [23]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    Recurrent Event Network: Autoregressive Structure Inferenceover Temporal Knowledge Graphs , author=. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    2020

  16. [24]

    Proceedings of the 31st ACM International Conference on Information & Knowledge Management , pages=

    Along the Time: Timeline-traced Embedding for Temporal Knowledge Graph Completion , author=. Proceedings of the 31st ACM International Conference on Information & Knowledge Management , pages=

  17. [25]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    HGE: embedding temporal knowledge graphs in a product space of heterogeneous geometric subspaces , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  18. [26]

    international conference on machine learning , pages=

    Know-evolve: Deep temporal reasoning for dynamic knowledge graphs , author=. international conference on machine learning , pages=. 2017 , organization=

    2017

  19. [27]

    Proceedings of the 28th International Conference on Computational Linguistics , pages=

    TeRo: A Time-aware Knowledge Graph Embedding via Temporal Rotation , author=. Proceedings of the 28th International Conference on Computational Linguistics , pages=

  20. [28]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Diachronic embedding for temporal knowledge graph completion , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  21. [29]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    ChronoR: Rotation Based Temporal Knowledge Graph Embedding , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  22. [30]

    Companion proceedings of the the web conference 2018 , pages=

    Deriving validity time in knowledge graph , author=. Companion proceedings of the the web conference 2018 , pages=

    2018

  23. [31]

    Communications of the ACM , volume=

    Maintaining knowledge about temporal intervals , author=. Communications of the ACM , volume=. 1983 , publisher=

    1983

  24. [32]

    Harvard Dataverse , volume=

    Icews event aggregations , author=. Harvard Dataverse , volume=

  25. [33]

    Learning Sequence Encoders for Temporal Knowledge Graph Completion

    Garc \'i a-Dur \'a n, Alberto and Duman c i \'c , Sebastijan and Niepert, Mathias. Learning Sequence Encoders for Temporal Knowledge Graph Completion. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018. doi:10.18653/v1/D18-1516

    2018 doi

  26. [34]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  27. [35]

    Advances in Neural Information Processing Systems , volume=

    Cape: Encoding relative positions with continuous augmented positional embeddings , author=. Advances in Neural Information Processing Systems , volume=

  28. [36]

    The Twelfth International Conference on Learning Representations , year=

    Towards Foundation Models for Knowledge Graph Reasoning , author=. The Twelfth International Conference on Learning Representations , year=

  29. [37]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    A Unified Temporal Knowledge Graph Reasoning Model Towards Interpolation and Extrapolation , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  30. [38]

    Proceedings of the 46th international ACM SIGIR conference on research and development in information retrieval , pages=

    Learn from relational correlations and periodic events for temporal knowledge graph reasoning , author=. Proceedings of the 46th international ACM SIGIR conference on research and development in information retrieval , pages=

  31. [39]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

    Complex Evolutional Pattern Learning for Temporal Knowledge Graph Reasoning , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

  32. [40]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    TimeTraveler: Reinforcement Learning for Temporal Knowledge Graph Forecasting , author=. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    2021

  33. [41]

    International conference on learning representations , year=

    Explainable subgraph reasoning for forecasting on temporal knowledge graphs , author=. International conference on learning representations , year=

  34. [42]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Tlogic: Temporal logical rules for explainable link forecasting on temporal knowledge graphs , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  35. [43]

    , author=

    TiRGN: Time-Guided Recurrent Graph Network with Local-Global Historical Patterns for Temporal Knowledge Graph Reasoning. , author=. IJCAI , pages=

  36. [44]

    Advances in neural information processing systems , volume=

    Neural bellman-ford networks: A general graph neural network framework for link prediction , author=. Advances in neural information processing systems , volume=

  37. [45]

    International conference on machine learning , pages=

    Inductive relation prediction by subgraph reasoning , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  38. [46]

    Advances in Neural Information Processing Systems , volume=

    Indigo: Gnn-based inductive knowledge graph completion using pair-wise encoding , author=. Advances in Neural Information Processing Systems , volume=

  39. [47]

    Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval , pages=

    Meta-knowledge transfer for inductive knowledge graph embedding , author=. Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval , pages=

  40. [48]

    International Conference on Machine Learning , pages=

    InGram: Inductive knowledge graph embedding via relation graphs , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  41. [49]

    Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

    Incorporating structured sentences with time-enhanced bert for fully-inductive temporal relation prediction , author=. Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

  42. [50]

    Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    ECOLA: Enhancing Temporal Knowledge Embeddings with Contextualized Language Representations , author=. Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    2023

  43. [51]

    Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    Pre-trained Language Model with Prompts for Temporal Knowledge Graph Completion , author=. Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    2023

  44. [52]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Temporal Knowledge Graph Forecasting Without Knowledge Using In-Context Learning , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  45. [53]

    2024 IEEE 40th International Conference on Data Engineering (ICDE) , pages=

    Local-global history-aware contrastive learning for temporal knowledge graph reasoning , author=. 2024 IEEE 40th International Conference on Data Engineering (ICDE) , pages=. 2024 , organization=

    2024

  46. [54]

    Joint European conference on machine learning and knowledge discovery in databases , pages=

    Comparing apples and oranges? on the evaluation of methods for temporal knowledge graph forecasting , author=. Joint European conference on machine learning and knowledge discovery in databases , pages=. 2023 , organization=

    2023

  47. [55]

    Proceedings of the 2021 conference on empirical methods in natural language processing , pages=

    Learning neural ordinary equations for forecasting future links on temporal knowledge graphs , author=. Proceedings of the 2021 conference on empirical methods in natural language processing , pages=

    2021

  48. [56]

    NAACL-HLT (Findings) , year=

    GenTKG: Generative Forecasting on Temporal Knowledge Graph with Large Language Models , author=. NAACL-HLT (Findings) , year=

  49. [57]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    zrLLM: Zero-Shot Relational Learning on Temporal Knowledge Graphs with Large Language Models , author=. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2024

  50. [58]

    The Third Learning on Graphs Conference , year=

    TRIX: A More Expressive Model for Zero-shot Domain Transfer in Knowledge Graphs , author=. The Third Learning on Graphs Conference , year=

  51. [59]

    Proceedings of the 30th ACM international conference on information & knowledge management , pages=

    Complex temporal question answering on knowledge graphs , author=. Proceedings of the 30th ACM international conference on information & knowledge management , pages=

  52. [60]

    NeurIPS , year =

    A Prompt-based Knowledge Graph Foundation Model for Universal In-Context Reasoning , author =. NeurIPS , year =

  53. [61]

    International conference on machine learning , pages=

    Complex embeddings for simple link prediction , author=. International conference on machine learning , pages=. 2016 , organization=

    2016

  54. [62]

    arXiv preprint arXiv:2401.06072 , year=

    Chain of history: Learning and forecasting with llms for temporal knowledge graph completion , author=. arXiv preprint arXiv:2401.06072 , year=

    arXiv

  55. [63]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.