Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

CognTKE: A Cognitive Temporal Knowledge Extrapolation Framework

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read CognTKE claims that combining global one-hop historical facts with recent local multi-hop paths in a query-specific layered digraph improves temporal knowledge graph extrapolation accuracy and yields interpretable, zero-shot-transferable…

desk verdict Solid incremental TKG extrapolation method with a real retrieval-ceiling flaw and an overclaimed abstract; worth a serious referee, but the WIKI numbers undercut 'significant improvement'. read the letter →

arxiv 2412.16557 v1 pith:CG2L6DTN submitted 2024-12-21 cs.AI

Wei Chen , Yuting Wu , Shuhan Wu , Zhiyu Zhang , Mengqi Liao , Youfang Lin , Huaiyu Wan This is my paper · ORCID
classification cs.AI
keywords temporalknowledgegraphextrapolationexplainablereasoningdualprocesstheoryattentionnetworkzero-shotrelationpathsTCR-Digraph
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

Future-event forecasting on temporal knowledge graphs usually follows either embedding-based models, which are accurate but opaque, or path-based models, which are explainable but restrict evidence to recent local time windows and lose longer historical relations. The paper tries to get the best of both by proposing CognTKE, which builds a query-specific layered digraph (the TCR-Digraph) whose first layer contains all global one-hop facts from the query entity over the whole past and whose later layers contain local multi-hop paths from the most recent 15 steps. Reasoning proceeds in two stages aligned with the cognitive science distinction between fast, intuitive one-hop judgment and slow, deliberate multi-hop deliberation. The paper's claim is that this design outperforms prior baselines on four standard datasets (largest gains on ICEWS18 and ICEWS05-15) and transfers to new datasets that share relations with the training data, while retaining an explicit chain of evidence for each prediction. A sympathetic reader would care because the framework addresses the explainability-accuracy trade-off that currently splits TKG extrapolation methods.

What carries the argument

The load-bearing object is the TCR-Digraph, a query-specific layered digraph whose first layer holds global one-hop historical facts from the whole past and whose later layers hold local multi-hop facts from the most recent m time steps (m=15). Encoding over it is carried by the TR-Component, which maps each relation plus its relative time to a vector via sinusoidal position encoding and an FFN, and the TR-GAT, which propagates messages along edges using a query-aware GRU variant (QTR GRU) that preserves the order of consecutive temporal relations and GAT-style attention that weights edges by relevance to the query. The global shallow reasoner and local deep reasoner apply this same encoder to layer 1 and layers 2..L, respectively, and update a candidate rule pool that serves as working memory.

What would settle it

Take a test query whose correct object is reachable only by a multi-hop relational chain that ends more than 15 time steps before the query and is not attached through any single-hop fact involving the query entity; if the framework still ranks the correct entity above decoys, the window restriction is not actually binding, and if it does not, the central design assumption fails.

Watch

Extended reading notes

Core claim

The central discovery proposed by the authors is that a temporal knowledge graph extrapolation model does not have to choose between global coverage and local specificity. Their TCR-Digraph keeps the first hop of every query in the full historical record, capturing long-range one-hop relations that path-based models miss, while bounding deeper chains to a recent window where multi-hop evidence is cheap and reliable. Encoding this digraph with a time-aware relation component and a query-aware graph attention network that preserves relation order, CognTKE learns to weight exactly those edges that support the query, which the paper formalizes as Theorem 1: any temporal logical rule that corresponds to a path in the TCR-Digraph can be encoded by some attention-parameter setting. Empirically the paper reports higher MRR and Hits@1/3/10 than the compared baselines on ICEWS14, ICEWS18, ICEWS05-15 and WIKI, with the clearest margins on the two larger ICEWS datasets, and reports zero-shot results where models trained on one ICEWS dataset are evaluated on another.

Load-bearing premise

The constructed graph for a query keeps only one-hop facts from the entire past but restricts multi-hop chains to the most recent 15 time steps, so any decisive chain that is both multi-hop and older than that window is invisible to the model.

Editorial extensions

If this is right

  • If the claim is right, path-based TKG extrapolation can recover long-range historical evidence without enumerating all multi-hop paths from the whole history, since the global one-hop layer preserves the long-range signal and the local layers refine it.
  • The two-reasoner design gives a concrete architectural meaning to System 1/System 2 reasoning, making the split between cheap global retrieval and expensive local deliberation a design choice for other graph sequence models.
  • The zero-shot results imply that models trained on one event graph can be applied to a later event graph sharing the same relation vocabulary, which matters for real deployments where labeled future data is unavailable.
  • Because the attention weights select the edges that explain a prediction, the extracted TCR-Digraph can serve as an explanation artifact for downstream users.

Reading between the lines

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

  • The paper's own parameter study shows performance rises slowly as the local window m grows, which suggests a testable extension: an adaptive window that widens when no high-confidence path is found would likely trade off the reported runtime against accuracy.
  • The TCR-Digraph construction creates a natural recursive refinement scheme: each layer is a subgraph expansion from the previous layer's candidate entities, so the same machinery could be applied to static knowledge graph reasoning or to inductive settings where entities are unseen at training time.
  • The 'candidate rule pool' and the QTR-GRU order encoding jointly predict that the model's sensitivity to relation order should increase with path depth; an experiment that shuffles temporal order within paths of length 2 vs length 4 would test that prediction directly.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 CognTKE, a temporal knowledge graph extrapolation framework built around a query-specific temporal cognitive relation digraph (TCR-Digraph). The digraph's first layer is populated with global one-hop historical facts involving the query entity over the entire history, while subsequent layers are populated with local multi-hop facts inside a recent time window [t_q-m, t_q] with m=15. A 'global shallow reasoner' (System 1) and a 'local deep reasoner' (System 2) recursively encode the digraph using a temporal-relation component and a graph attention network with a QTR-GRU message-passing variant, after which an MLP decoder scores entities. Experiments on ICE14, ICE18, ICE05-15 and WIKI compare against several embedding- and path-based baselines, with ablations, zero-shot transfer across ICEWS datasets, parameter studies and a prediction-time comparison. The paper claims significant accuracy improvements and strong zero-shot and interpretability abilities.

Significance. If the results held as stated, the contribution would be a useful hybrid of path-based explainability and subgraph-based reasoning: the TCR-Digraph provides rule-like evidence, and the two-level reasoner gives an interpretable attention structure. The code is open-sourced and the evaluation is on standard benchmarks, which strengthens reproducibility. However, the current evidence is not sufficient to establish the headline claims: the retrieval construction imposes a structural recall ceiling that is not measured, the theoretical guarantee is a conditional existence statement rather than a property of the trained model, and the experimental numbers are mixed across datasets with no uncertainty quantification. The framework is plausible and worth revising, but the significance claimed in the abstract is not yet supported.

major comments (4)
  1. [Algorithm 1 (Appendix A.2) and Eq. (10)] The TCR-Digraph construction is a hard retrieval ceiling. The first layer samples global one-hop facts from the whole history, but every subsequent layer samples only from {G_{tq-m},...,G_{tq-1}} with m=15 (Algorithm 1, lines 9–11). Equation (10) then assigns score 0 to every entity not in the digraph. Consequently, any ground-truth object whose only evidential path has length at least 2 and whose non-first edges occur before tq-15 is structurally unpredicatable, regardless of how good the reasoner is. The paper varies m in Figure 6 but never measures the fraction of test answers that are present in the assembled digraph ('graph recall'), which is the actual ceiling. This is especially relevant on WIKI, where Hits@10 (87.34) is far below TempValid (97.54); the deficit could be a retrieval cap rather than a ranking defect. Please report graph-recall statistics and, if the claim is that longer historical paths are preserved, extend the window for multi-hop layers or restrict the claim accordingly.
  2. [Appendix C (Proof of Theorem 1)] Theorem 1 and its proof assume the desired behavior rather than deriving it. The theorem states that if there exists a parameter setting Θ and threshold λ with attention weights above θ exactly on the desired edges, then the target graph can be recovered; Appendix C then uses the universal approximation theorem to claim such parameters exist for each layer independently. This does not show that the trained model encodes the rule, and the per-layer independence of the constructed parameters is inconsistent with the shared TR-GAT and QTR-GRU weights in Eqs. (4)–(9). As written, the theorem is a near-tautological conditional; it does not justify the sentence in 'Interpretable Analysis' that CognTKE 'is capable of encoding any temporal logical rule that corresponds to a path in the TCR-Digraph.' Please either provide a genuinely constructive proof of an approximation statement, or explicitly mark the theorem as an idealized existence result with no guarantee from training.
  3. [Table 1 and Abstract] The headline claim of 'significant improvement' is contradicted by Table 1 on two of four datasets. On WIKI, CognTKE has MRR 83.21 vs. TempValid 83.19 and Hits@10 87.34 vs. 97.54; on ICE14, Hits@10 is 64.49 vs. 65.06. No error bars, confidence intervals, or significance tests are reported anywhere in the paper, so 'significant' is not established. Please report multiple seeds or equivalent uncertainty measures and either temper the abstract or demonstrate significance on the metrics where the improvement is real, such as ICE18 MRR (35.24 vs. 33.50).
  4. [Table 3 (Zero-Shot Reasoning)] The zero-shot results provide only partial support for 'excellent zero-shot reasoning ability.' When ICE05-15 is the test set, CognTKE trained on ICE14 (MRR 49.99) or on ICE18 (MRR 49.86) is slightly below TempValid trained on ICE05-15 (MRR 50.31), and the gains on other transfer directions are modest. Please report zero-shot comparisons with uncertainty estimates and state precisely which transfer configurations support the claim.
minor comments (5)
  1. [Figure 3 caption] The caption contains the typo 'TCR-Digrph'; please correct it.
  2. [Eq. (3)] The dimensions of W1 and W2 appear inconsistent with the concatenation [h_r, v̄_t]; if h_r ∈ R^d and v̄_t ∈ R^{d_time}, W1 should map R^{d+d_time} to R^d, and W2 should map R^d to R^d. Please clarify.
  3. [Algorithm 1] The pseudocode uses F_global and F_local without defining them, and the loop 'while t < |T|' is not tied to the query time t_q; please align the pseudocode with the description in the 'Temporal Cognitive Relation Digraph' section.
  4. [Ablation Study] The ablation paragraph begins with 'As reported in Table 2.-w/o System1...'; the period after 'Table 2' should be a comma or the sentence should be rephrased.
  5. [Overall Results] In the comparison with path-based baselines, 'TLogic, TlTer' should read 'TLogic, TITer'.

Circularity Check

1 steps flagged · score 3.0 of 10

Main benchmark results are externally evaluated and not circular; the only near-circular element is Theorem 1's existential proof, which assumes the very parameter setting that the conclusion asserts, plus minor non-load-bearing self-citations.

  1. self definitional [Interpretable Analysis (Theorem 1) and Appendix C (Proofs for Theorem 1)]
    "Assuming there exists a directed graph ĜC that is built by C, a parameter setting Θ, and a threshold λ ∈ (0, 1) for CognTKE. ĜC can equal to the TCR-digraph if CognTKE’s edges have attention weights cℓ es,rt|r tq q > θ in Ĝ..., where θ is a learned decision boundary parameter. ... According to the universal approximation theorem, there exists a set of parameters ... that can learn a decision boundary θ ... if c > θ and otherwise not."

    The theorem’s conclusion — that CognTKE is capable of encoding any temporal logical rule corresponding to a path in the TCR-Digraph — is obtained by assuming, layer by layer, that a parameter setting exists making the attention weight c exceed threshold θ exactly on the edges of the target graph ĜC. That existential assumption is precisely the encoding capability the theorem claims to prove; the universal-approximation invocation restates the same assumption in general form rather than deriving it from the loss or training dynamics. Consequently, “ĜC can equal the TCR-Digraph” is true by construction whenever the assumed parameters happen to exist, making the interpretability guarantee near-tautological.

full rationale

Score is 3.0 rather than 0 because of the near-tautological Theorem 1 and minor non-load-bearing self-citations (Chen et al. 2022b; Chen et al. 2024; Wu et al. 2023), but the central empirical claims are evaluated on four public benchmarks against external baselines, so no fitted parameter is renamed as a prediction. The local-window construction (Appendix B.2, Algorithm 1, m=15) imposes a retrieval ceiling, but that is a coverage/correctness limitation, not circularity. No self-citation chain forces the main result, and the zero-shot and ablation studies are also externally grounded rather than defined by the model's own outputs.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the chosen hyperparameters, the sufficiency of the graph retrieval strategy, and a non-constructive expressiveness argument. These are architectural and domain assumptions rather than new physical entities, but they still shape what the model can and cannot do.

free parameters (6)
  • local time window m = 15
    Window length for retrieving local multi-hop temporal paths; chosen by hand and analyzed in Appendix B.
  • number of TR-GAT layers L = 4
    Depth of local reasoning; selected based on performance plateau (Figure 7).
  • relation embedding size d = 64
    Embedding dimension; shown to have minimal impact (Figure 8).
  • time embedding size dtime = 32
    Time embedding dimension; minimal sensitivity (Figure 9).
  • learning rate = 0.001
    Adam learning rate.
  • batch size = 128
    Training batch size.
assumptions (4)
  • domain assumption The TCR-Digraph sampling from global history (layer 1) and recent window (layers 2+) preserves all evidence needed for accurate extrapolation.
    Central to the method; if decisive evidence lies outside the window or is not retrieved, the prediction will miss it.
  • standard math The universal approximation theorem applies to the specific attention-based architecture used in CognTKE.
    Used in Appendix C to assert existence of parameters that make attention weights exceed a threshold; this is a non-constructive existence claim.
  • domain assumption Temporal relations can be represented as a combination of relation embedding and sinusoidal time embedding via Eq. 3.
    The TR-Component assumes this decomposition captures both semantic and temporal aspects.
  • domain assumption Inverse and identity relations are sufficient to model queries of both subject and object prediction.
    Standard preprocessing in TKG reasoning; the model relies on these augmented triples.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CognTKE: A Cognitive Temporal Knowledge Extrapolation Framework." pith.science (2026). https://pith.science/paper/CG2L6DTN

@misc{pith2026241216557,
  author       = {Pith},
  title        = {Pith review of: CognTKE: A Cognitive Temporal Knowledge Extrapolation Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CG2L6DTN}},
  note         = {Machine review of arXiv:2412.16557}
}
read the original abstract

Reasoning future unknowable facts on temporal knowledge graphs (TKGs) is a challenging task, holding significant academic and practical values for various fields. Existing studies exploring explainable reasoning concentrate on modeling comprehensible temporal paths relevant to the query. Yet, these path-based methods primarily focus on local temporal paths appearing in recent times, failing to capture the complex temporal paths in TKG and resulting in the loss of longer historical relations related to the query. Motivated by the Dual Process Theory in cognitive science, we propose a \textbf{Cogn}itive \textbf{T}emporal \textbf{K}nowledge \textbf{E}xtrapolation framework (CognTKE), which introduces a novel temporal cognitive relation directed graph (TCR-Digraph) and performs interpretable global shallow reasoning and local deep reasoning over the TCR-Digraph. Specifically, the proposed TCR-Digraph is constituted by retrieving significant local and global historical temporal relation paths associated with the query. In addition, CognTKE presents the global shallow reasoner and the local deep reasoner to perform global one-hop temporal relation reasoning (System 1) and local complex multi-hop path reasoning (System 2) over the TCR-Digraph, respectively. The experimental results on four benchmark datasets demonstrate that CognTKE achieves significant improvement in accuracy compared to the state-of-the-art baselines and delivers excellent zero-shot reasoning ability. \textit{The code is available at https://github.com/WeiChen3690/CognTKE}.

Figures

Figures reproduced from arXiv: 2412.16557 by the authors.

Figure 1
Figure 1. An example illustrates longer history temporal [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An illustrative diagram of the proposed CognTKE architecture. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An illustrative of TCR-Digrph Gˆ e1,e3|3 formed by the TKG. In practice, such TR-Digraph can bring much earlier his￾torical information that is not directly relevant to the query, which not only makes the TR-Digraph construction and en￾coding expensive but also does not facilitate subsequent en￾tity reasoning. Therefore, we propose an improved TCR￾Digraph Gˆ e tq q ,et o |L based on TR-Digraph, which contains im￾por… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Visualization of the learned structures. Dashed [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The Component of QTR GRU. A.2 Pseudocode. The algorithm for the TCR-Digraph en￾coding in CognTKE is described in Algorithm 1. 1 3 5 7 9 11 13 15 m 40 42 44 46 48 50 52 54 MRR Hits@3 (a) ICE14 1 3 5 7 9 11 13 15 m 46 48 50 52 54 56 58 60 MRR Hits@3 (b) ICE05-15 [PITH_F…
Figure 6
Figure 6. Figure 6: Study on the length of the local time window [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 8
Figure 8. Figure 8: Study on different embedding size of relation [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Study on different embedding size of time [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A Generative Adaptive Replay Continual Learning Model for Temporal Knowledge Graph Reasoning

    cs.IR 2025-06 conditional novelty 6.0 of 10

    DGAR uses diffusion-generated, model-guided historical entity distributions with adaptive replay to reduce catastrophic forgetting in temporal knowledge graph reasoning.

Reference graph

Works this paper leans on

39 extracted references · 36 canonical work pages · cited by 1 Pith paper

  1. [1]

    Baddeley, A. 1992. Working memory. Science, 255(5044): 556--559

    work page 1992

  2. [2]

    D.; Eades, P.; Tamassia, R.; and Tollis, I

    Battista, G. D.; Eades, P.; Tamassia, R.; and Tollis, I. G. 1999. Graph Drawing: Algorithms for the Visualization of Graphs. Prentice-Hall. ISBN 0-13-301615-3

    work page 1999

  3. [3]

    Chen, K.; Wang, Y.; Li, Y.; and Li, A. 2022 a . RotateQVS: Representing Temporal Information as Rotations in Quaternion Vector Space for Temporal Knowledge Graph Completion. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022 , 5843--5857. Associatio...

    work page 2022

  4. [4]

    Chen, W.; Wan, H.; Guo, S.; Huang, H.; Zheng, S.; Li, J.; Lin, S.; and Lin, Y. 2022 b . Building and exploiting spatial-temporal knowledge graph for next POI recommendation. Knowl. Based Syst., 258: 109951

    work page 2022

  5. [5]

    Chen, W.; Wan, H.; Wu, Y.; Zhao, S.; Cheng, J.; Li, Y.; and Lin, Y. 2024. Local-Global History-Aware Contrastive Learning for Temporal Knowledge Graph Reasoning. In 40th IEEE International Conference on Data Engineering, ICDE 2024, Utrecht, The Netherlands, May 13-16, 2024 , 733--746. IEEE

    work page 2024

  6. [6]

    Cho, K.; van Merrienboer, B.; G \" u l c ehre, C .; Bahdanau, D.; Bougares, F.; Schwenk, H.; and Bengio, Y. 2014. Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, 1724--1734. ACL

    work page 2014

  7. [7]

    Evans; and BT, J. S. 1984. Heuristic and analytic processes in reasoning. British Journal of Psychology, 75(4): 451--468

    work page 1984

  8. [8]

    Evans, J. S. B. 2003. In Two Minds: Dual-process Accounts of Reasoning. Trends in cognitive sciences, 7(10): 454--459

    work page 2003

Show all 39 references
  1. [9]

    Evans, J. S. B. 2008. Dual-processing Accounts of Reasoning, Judgment, and Social Cognition. Annu. Rev. Psychol., 59: 255--278

    2008

  2. [10]

    Guan, S.; Cheng, X.; Bai, L.; Zhang, F.; Li, Z.; Zeng, Y.; Jin, X.; and Guo, J. 2023. What is Event Knowledge Graph: A Survey. IEEE Trans. Knowl. Data Eng. , 35(7): 7569--7589

    2023

  3. [11]

    Han, Z.; Ding, Z.; Ma, Y.; Gu, Y.; and Tresp, V. 2021. Learning Neural Ordinary Equations for Forecasting Future Links on Temporal Knowledge Graphs. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 8352--8364. Association for Computati...

    2021

  4. [12]

    Huang, R.; Wei, W.; Qu, X.; Zhang, S.; Chen, D.; and Cheng, Y. 2024. Confidence is not Timeless: Modeling Temporal Validity for Rule-based Temporal Knowledge Graph Forecasting. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1...

    2024

  5. [13]

    Jin, W.; Qu, M.; Jin, X.; and Ren, X. 2020. Recurrent Event Network: Autoregressive Structure Inferenceover Temporal Knowledge Graphs. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, 6669--6683. Association for Computational Linguistics

    2020

  6. [14]

    Lacroix, T.; Usunier, N.; and Obozinski, G. 2018. Canonical Tensor Decomposition for Knowledge Base Completion. In Proceedings of the 35th International Conference on Machine Learning, ICML 2018, Stockholmsm \" a ssan, Stockholm, Sweden, July 10-15, 2018 , volume 80 of Proceed...

    2018

  7. [15]

    Leblay, J.; and Chekol, M. W. 2018. Deriving Validity Time in Knowledge Graph. In Champin, P.; Gandon, F.; Lalmas, M.; and Ipeirotis, P. G., eds., The-Web-Conf, 1771--1776. ACM

    2018

  8. [16]

    Li, Y.; Sun, S.; and Zhao, J. 2022. Tirgn: time-guided recurrent graph network with local-global historical patterns for temporal knowledge graph reasoning. In IJCAI, 2152--2158

    2022

  9. [17]

    Li, Z.; Guan, S.; Jin, X.; Peng, W.; Lyu, Y.; Zhu, Y.; Bai, L.; Li, W.; Guo, J.; and Cheng, X. 2022. Complex Evolutional Pattern Learning for Temporal Knowledge Graph Reasoning. In ACL, 290--296. Association for Computational Linguistics

    2022

  10. [18]

    Li, Z.; Jin, X.; Guan, S.; Li, W.; Guo, J.; Wang, Y.; and Cheng, X. 2021 a . Search from History and Reason for Future: Two-stage Reasoning on Temporal Knowledge Graphs. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th Int...

    2021

  11. [19]

    Li, Z.; Jin, X.; Li, W.; Guan, S.; Guo, J.; Shen, H.; Wang, Y.; and Cheng, X. 2021 b . Temporal Knowledge Graph Reasoning Based on Evolutional Representation Learning. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retr...

    2021

  12. [20]

    Liang, K.; Meng, L.; Liu, M.; Liu, Y.; Tu, W.; Wang, S.; Zhou, S.; Liu, X.; Sun, F.; and He, K. 2024 a . A Survey of Knowledge Graph Reasoning on Graph Types: Static, Dynamic, and Multi-modal. IEEE Transactions on Pattern Analysis and Machine Intelligence

    2024

  13. [21]

    Liang, K.; Meng, L.; Zhou, S.; Tu, W.; Wang, S.; Liu, Y.; Liu, M.; Zhao, L.; Dong, X.; and Liu, X. 2024 b . MINES: Message Intercommunication for Inductive Relation Reasoning over Neighbor-Enhanced Subgraphs. In Proceedings of the AAAI Conference on Artificial Intelligence, vo...

    2024

  14. [22]

    Lin, Q.; Liu, J.; Mao, R.; Xu, F.; and Cambria, E. 2023. TECHS: Temporal logical graph networks for explainable extrapolation reasoning. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 1281--1293

    2023

  15. [23]

    Liu, K.; Zhao, F.; Chen, H.; Li, Y.; Xu, G.; and Jin, H. 2022 a . DA-Net: Distributed Attention Network for Temporal Knowledge Graph Reasoning. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, Atlanta, GA, USA, October 17-21, 2022 ...

    2022

  16. [24]

    Liu, K.; Zhao, F.; Xu, G.; Wang, X.; and Jin, H. 2023. RETIA: Relation-Entity Twin-Interact Aggregation for Temporal Knowledge Graph Extrapolation. In 39th IEEE International Conference on Data Engineering, ICDE 2023, Anaheim, CA, USA, April 3-7, 2023 , 1761--1774. IEEE

    2023

  17. [25]

    Liu, Y.; Ma, Y.; Hildebrandt, M.; Joblin, M.; and Tresp, V. 2022 b . TLogic: Temporal Logical Rules for Explainable Link Forecasting on Temporal Knowledge Graphs. In Proceedings of the AAAI conference on artificial intelligence, 4120--4127. AAAI Press

    2022

  18. [26]

    Mei, X.; Yang, L.; Cai, X.; and Jiang, Z. 2022. An Adaptive Logical Rule Embedding Model for Inductive Reasoning over Temporal Knowledge Graphs. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 7304--7316. Association for Computational...

    2022

  19. [27]

    Pan, J.; Nayyeri, M.; Li, Y.; and Staab, S. 2024. HGE: Embedding Temporal Knowledge Graphs in a Product Space of Heterogeneous Geometric Subspaces. In Thirty-Eighth AAAI Conference on Artificial Intelligence, February 20-27, 2024, Vancouver, Canada , 8913--8920. AAAI Press

    2024

  20. [28]

    Sloman; and A, S. 1996. The empirical case for two systems of reasoning. Psychological bulletin, 119(1): 3

    1996

  21. [29]

    Sun, H.; Zhong, J.; Ma, Y.; Han, Z.; and He, K. 2021. TimeTraveler : Reinforcement Learning for Temporal Knowledge Graph Forecasting. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 8306--8319. Association for Computational Linguistics

    2021

  22. [30]

    Timoth \' e e, L.; Obozinski, G.; and Usunier, N. 2020. Tensor Decompositions for Temporal Knowledge Base Completion. In International Conference on Learning Representations. OpenReview.net

    2020

  23. [31]

    N.; Kaiser, L.; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L.; and Polosukhin, I. 2017. Attention is All you Need. In NeurIPS, 5998--6008

    2017

  24. [32]

    Velickovic, P.; Cucurull, G.; Casanova, A.; Romero, A.; Li \` o , P.; and Bengio, Y. 2018. Graph Attention Networks. In International Conference on Learning Representations. OpenReview.net

    2018

  25. [33]

    Wu, S.; Wan, H.; Chen, W.; Wu, Y.; Shen, J.; and Lin, Y. 2023. Towards Enhancing Relational Rules for Knowledge Graph Link Prediction. In Findings of the Association for Computational Linguistics: EMNLP 2023, Singapore, December 6-10, 2023 , 10082--10097. Association for Compu...

    2023

  26. [34]

    Xin, H.; and Chen, L. 2024. KartGPS: Knowledge Base Update with Temporal Graph Pattern-based Semantic Rules. In 40th IEEE International Conference on Data Engineering, ICDE 2024, Utrecht, The Netherlands, May 13-16, 2024 , 5075--5087. IEEE

    2024

  27. [35]

    Xu, Y.; Ou, J.; Xu, H.; and Fu, L. 2023. Temporal knowledge graph reasoning with historical contrastive learning. In Proceedings of the AAAI conference on artificial intelligence, volume 37, 4765--4773

    2023

  28. [36]

    Zhang, Y.; and Yao, Q. 2022. Knowledge Graph Reasoning with Relational Digraph. In The-Web-Conf, 912--924. ACM

    2022

  29. [37]

    Zhu, C.; Chen, M.; Fan, C.; Cheng, G.; and Zhang, Y. 2021. Learning from History: Modeling Temporal Knowledge Graphs with Sequential Copy-Generation Networks. In Proceedings of the AAAI conference on artificial intelligence, 4732--4740. AAAI Press

    2021

  30. [38]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  31. [39]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.