Pith. sign in

REVIEW 3 major objections 6 minor 43 references

T3DM: Test-Time Training-Guided Distribution Shift Modelling for Temporal Knowledge Graph Reasoning

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A plug-in training loop that adapts temporal knowledge graph models to shifting event distributions at test time improves most link-prediction metrics across five benchmark datasets.

desk verdict The paper's central GAN objective contradicts its own prose—the generator is rewarded for easy negatives—so the empirical results do not support the advertised method. read the letter →

arxiv 2507.01597 v1 pith:4277EMYA submitted 2025-07-02 cs.AI cs.CL

Yuehang Si , Zefan Zeng , Jincai Huang , Qing Cheng This is my paper · ORCID
classification cs.AIcs.CL
keywords temporalknowledgegraphreasoningtest-timetrainingdistributionshiftnegativesamplinggenerativeadversarialnetworkslinkpredictionLSTM
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

Temporal knowledge graphs record facts that change over time, and reasoning models must predict future facts despite the fact that the mix of event types shifts between training and test periods. This paper proposes T3DM, a plug-and-play training procedure that attacks that shift in two ways. At test time, an LSTM forecasts the next timestamp's entity-count distribution, and the link-prediction model is further trained to match that forecast, adapting to the new distribution. During training, an adversarial generator (TKGAN) produces harder negative samples than random entity substitution, reducing the 'zero-loss' problem that makes training signals uninformative. The paper reports that wrapping five baseline TKG models with T3DM improves most link-prediction metrics on five datasets, and that the test-time component alone improves CyGNet and Co-CENET on all metrics across the five datasets.

What carries the argument

The load-bearing object is a two-part training wrapper around an existing TKG link-prediction model. The first part is a test-time auxiliary task: an LSTM (a long short-term memory neural network) is trained on the historical sequence of entity-count distributions, and at test time it forecasts the distribution for the next timestamp; the target model's own predicted distribution is aligned to this forecast by a cross-entropy loss, so the model continues to learn during inference. The second part is TKGAN, a GAN-style negative sampler: a generator scores candidate negative quadruples and samples hard negatives, while the discriminator (or the target model) scores positive and negative quadruples, and the generator is updated by a policy-gradient reward with a baseline to reduce variance. These two parts are what carry the claimed gains: the auxiliary task adapts the model to shifted event distributions, and the adversarial sampler supplies training signal in place of uninformative random negatives.

What would settle it

Use the true future entity-count distribution in the test-time auxiliary loss instead of the LSTM's forecast and compare against T3DM; if the true distribution does not beat the no-adaptation baseline, the distribution-forecast task is not the mechanism driving the reported gains.

Watch

Extended reading notes

Core claim

The paper's central claim is that a temporal knowledge graph reasoning model can be made more accurate and more robust by explicitly modelling event distribution shift rather than only modelling repeated global facts and local historical patterns. The proposed framework, T3DM, treats the entity-count distribution at each timestamp as a measurable proxy for the event-type distribution, trains an LSTM to predict the next distribution, and then uses the cross-entropy between the LSTM's forecast and the model's own predicted distribution as an auxiliary test-time loss that continues to update the model. Alongside this, TKGAN generates negative quadruples adversarially, with a generator trained by policy-gradient reinforcement learning to produce samples that the discriminator finds hard to distinguish from true facts. The experiments show that, in most configurations, baselines equipped with T3DM outperform the same baselines trained with random negative sampling, with complete gains on TATransE and DE-SimplE and only occasional slight metric drops elsewhere; the TTT component also improves CyGNet and Co-CENET on almost every metric and dataset.

Load-bearing premise

The method works only if the count of how often each entity appears at a moment is a faithful stand-in for the shift in event types that matters for prediction, and if the forecast of those counts is accurate enough to teach the model.

Editorial extensions

If this is right

  • Existing TKG reasoning models can be upgraded without architectural changes by adding T3DM as a plug-and-play training method.
  • Test-time adaptation through an auxiliary distribution-forecast task makes link prediction more robust to the event-type shifts that appear when moving from training to test timestamps.
  • Adversarial negative sampling reduces the zero-loss problem, so training provides more informative gradients than random entity substitution.
  • The test-time component alone, without TKGAN, improves models such as CyGNet and Co-CENET on nearly all metrics, suggesting the adaptation signal is useful independently.
  • Short LSTM lookback windows (around 20 timestamps) give the best results in the reported experiments, while longer windows degrade performance.

Reading between the lines

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

  • Because the entity-count distribution is a coarse aggregate, a finer auxiliary task involving per-relation or per-entity-type distributions might capture the shift that matters for link prediction even more directly; the paper does not test this variant.
  • The paper never reports the accuracy of the LSTM's distribution forecasts, so some of the gain attributed to distribution-shift modelling could instead come from the extra gradient updates acting as a regulariser.
  • TKGAN's reliance on the discriminator's score as a reward means the quality of generated negatives is bounded by the discriminator's own ranking ability; models with very different scoring functions may need a similarly matched generator to avoid occasional degradations.
  • The same test-time training recipe could plausibly transfer to other sequence-based reasoning tasks, such as temporal link prediction in evolving social networks, though that transfer is not demonstrated here.
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 / 6 minor

Summary. The paper proposes T3DM, a training framework for temporal knowledge graph reasoning that combines test-time training with an auxiliary LSTM task predicting future entity-count distributions, and TKGAN, a policy-gradient adversarial negative sampler. The method is applied to several TKGR baselines and evaluated with MRR and Hits@K on GDELT, ICEWS14, ICEWS18, YAGO11K, and Wikidata12K, with additional ablation and sensitivity experiments. The central claim is that T3DM yields more accurate and robust link prediction by adapting to event distribution shift at test time and by generating harder negative samples during training.

Significance. The high-level ideas are timely and relevant: distribution shift is a known issue in temporal knowledge graphs, and negative sampling quality directly affects embedding training. The paper ships an anonymous code release and evaluates across five datasets and multiple model families, which is a strength. If the method is sound and reproducible, the TTT auxiliary task is a novel contribution to TKGR. However, the current manuscript does not establish the method's validity: the auxiliary loss is not fully specified, the pseudo-labeling mechanism is not validated, and the experimental reporting lacks the statistical grounding needed to support the claimed gains. The significance of the contribution is therefore currently uncertain.

major comments (3)
  1. [§3.4, Eq. (7)] The auxiliary test-time loss is not computable as written. X_Pred, the "distribution predicted by the TKGR model," is never defined; Eq. (8) in §3.5 defines inference as an argmax over candidate entities, not a distribution over entities. The paper must specify how the model's predicted entity-count distribution is constructed from link prediction scores (e.g., by aggregating softmax outputs over a test batch) and how the loss is back-propagated through the target model. In addition, the LSTM-predicted pseudo-labels are used at test time without reporting the LSTM's forecasting accuracy or analyzing cases in which the auxiliary loss could push the model away from good solutions; the Limitations section acknowledges this gap, but it is load-bearing for the TTT contribution.
  2. [Table 1, §4.2] The experimental evidence does not support the textual claims. The paper states that "only one metric is slightly inferior to the baseline on TTransE and HyTE," but on the YAGO11K dataset TTransE's Hits@1 drops from 34.29 to 27.66 with T3DM, which is neither slight nor a single isolated case (TADistmult's Hits@1 drops from 62.80 to 59.74). No error bars, multiple seeds, or statistical tests are reported, so the small improvements in most rows (e.g., DE-SimplE MRR from 12.56 to 13.06 on ICEWS18) cannot be distinguished from optimization noise. The paper should report variance and significance, and correct the summary of Table 1.
  3. [§3.3, Algorithm 1] The policy-gradient derivation in Eq. (3) assumes negative quadruples are sampled from pG, but Algorithm 1 line 9 selects the candidate with the highest probability ph rather than sampling from pG. The update in line 12 then uses log ph of a deterministically chosen action, which is not an unbiased estimator of the R_G gradient in Eq. (2). The authors should either implement the sampling as derived or provide a theoretical justification for the greedy selection rule, since the RL-based adversarial sampling is a central contribution.
minor comments (6)
  1. [§3.3] The term "score" is used inconsistently. Eq. (1) treats f_D as a distance in a margin loss, while the prose says the generator should "maximise the score given by the discriminator." This apparent contradiction disappears if f_D is explicitly described as a distance and the prose says "maximise the negative distance" (as in Eq. (2)). Please clarify to avoid the misleading reading that the generator maximizes a plausibility score.
  2. [Eq. (7)] The cross-entropy expression has a formatting error: the sum lacks a closing parenthesis, and the definition of X_Pred is missing; please rewrite the equation with all variables defined.
  3. [Appendix A.2, Eq. (9)] The MRR formula uses a confusing factor 1/(2·N(G)) and sums over "o,s in G" without defining the query pairs; please make the ranking aggregation precise.
  4. [Figure 4] Figure 4 is described only in one sentence; the axes, the protocol for time-aware negative sampling, and whether the results are averaged over multiple runs are not specified.
  5. [Section 1, contribution 2] The claim of being the first to introduce TTT into the knowledge graph domain is too broad given that the related work already discusses TTT for graphs and recommender systems; please qualify the claim to temporal knowledge graph reasoning or comparable settings.
  6. [Section 4.4] The sensitivity analysis for LSTM input length reports only a figure without numerical values or standard deviations; report the actual metric values and the range of l tested.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central T3DM and TKGAN claims rest on independent empirical evaluation; the sole same-author citation is not load-bearing.

full rationale

The central derivation is not circular. The T3DM test-time update uses LSTM-predicted entity-count distributions as pseudo-labels for an auxiliary task; the LSTM is trained separately on historical entity-count sequences, and the final link-prediction metrics are evaluated against true future facts. The auxiliary loss therefore does not inject the test labels, and the paper's Limitations section explicitly acknowledges that test-phase labels are only LSTM predictions. The TKGAN adversarial sampler is described with a sign inconsistency, but that is an internal implementation/correctness issue, not a self-referential reduction of the claimed result to its inputs. The only same-author citation (Coherence-Mode) appears in related work and as a baseline in Table 2; it is not used to justify the central T3DM or TKGAN mechanisms, and the main results in Table 1 are against independent baselines. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported, and no existing result is repackaged under new coordinates. The paper is self-contained with respect to its empirical claims, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method introduces no new entities; it defines a training procedure. The main free parameters are hyperparameters (LSTM sequence length, margin, number of negatives, learning rate) and the unweighted auxiliary loss.

free parameters (5)
  • LSTM input sequence length l = 20
    A hyperparameter; the sensitivity analysis in Section 4.4 shows performance is best near length 20.
  • Margin gamma in discriminator loss = not reported
    Appears in Eq. (1); the paper does not state the value used in experiments.
  • Number of candidate negative quadruples K = not reported
    Algorithm 1 samples K negative candidates per positive; K is not specified.
  • Learning rate = selected from {0.0001, 0.0005, 0.001, 0.01}
    Chosen by tuning on validation data, as stated in Appendix A.1.
  • Weight of auxiliary test-time loss = not reported
    Section 3.4 does not describe a weighting factor; the scale of L_cmp relative to the main loss is unspecified.
assumptions (4)
  • domain assumption Entity-count distribution shift is a valid proxy for event-type distribution shift.
    Section 3.4 assumes that the distribution of entity counts can adequately reflect changes in event types.
  • ad hoc to paper LSTM-predicted future entity distributions can serve as reliable pseudo-labels for test-time adaptation.
    Section 3.4 uses LSTM outputs as pseudo-labels; no validation of forecasting accuracy is provided.
  • standard math The policy gradient theorem applies to the generator's discrete sampling.
    Eq. (3) uses the REINFORCE estimator, a standard result in reinforcement learning.
  • domain assumption Generated negative quadruples are almost surely true negatives because the candidate set is small and sampled uniformly.
    Section 3.3 argues that it is improbable that uniformly sampled candidates include false negatives, but this is not quantified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of T3DM: Test-Time Training-Guided Distribution Shift Modelling for Temporal Knowledge Graph Reasoning." pith.science (2026). https://pith.science/paper/4277EMYA

@misc{pith2026250701597,
  author       = {Pith},
  title        = {Pith review of: T3DM: Test-Time Training-Guided Distribution Shift Modelling for Temporal Knowledge Graph Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4277EMYA}},
  note         = {Machine review of arXiv:2507.01597}
}
read the original abstract

Temporal Knowledge Graph (TKG) is an efficient method for describing the dynamic development of facts along a timeline. Most research on TKG reasoning (TKGR) focuses on modelling the repetition of global facts and designing patterns of local historical facts. However, they face two significant challenges: inadequate modeling of the event distribution shift between training and test samples, and reliance on random entity substitution for generating negative samples, which often results in low-quality sampling. To this end, we propose a novel distributional feature modeling approach for training TKGR models, Test-Time Training-guided Distribution shift Modelling (T3DM), to adjust the model based on distribution shift and ensure the global consistency of model reasoning. In addition, we design a negative-sampling strategy to generate higher-quality negative quadruples based on adversarial training. Extensive experiments show that T3DM provides better and more robust results than the state-of-the-art baselines in most cases.

Figures

Figures reproduced from arXiv: 2507.01597 by the authors.

Figure 1
Figure 1. Illustration of the change of events in the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of event distribution shift and low [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The training framework of T3DM. The first part represents Training Time Training, and the second part [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison results of different negative sam [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Sensitivity analysis results of input sequence [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 37 canonical work pages

  1. [1]

    Antoine Bordes, Nicolas Usunier, Alberto Garc \' a - Dur \' a n, Jason Weston, and Oksana Yakhnenko. 2013. Translating embeddings for modeling multi-relational data. In NIPS 2013, pages 2787--2795

    work page 2013

  2. [2]

    Elizabeth Boschee, Jennifer Lautenschlager, Sean O’Brien, Steve Shellman, James Starz, and Michael Ward. 2015. Icews coded event data. Harvard Dataverse, 12

    work page 2015

  3. [3]

    Liwei Cai and William Yang Wang. 2018. KBGAN: adversarial learning for knowledge graph embeddings. In NAACL-HLT 2018, pages 1470--1480. Association for Computational Linguistics

    work page 2018

  4. [4]

    Shiming Chen, Wenjie Wang, Beihao Xia, Xinge You, Qinmu Peng, Zehong Cao, and Weiping Ding. 2021. CDE-GAN: cooperative dual evolution-based generative adversarial network. IEEE Trans. Evol. Comput. , 25(5):986--1000

    work page 2021

  5. [5]

    Zhongwu Chen, Chengjin Xu, Fenglong Su, Zhen Huang, and Yong Dou. 2023. Temporal extrapolation and knowledge transfer for lifelong temporal knowledge graph reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2023, Singapore, December 6-10, 2023 , pages 6736--6746. Association for Computational Linguistics

    work page 2023

  6. [6]

    Talukdar

    Shib Sankar Dasgupta, Swayambhu Nath Ray, and Partha P. Talukdar. 2018. Hyte: Hyperplane-based temporally aware knowledge graph embedding. In EMNLP 2021, pages 2001--2011. Association for Computational Linguistics

    work page 2018

  7. [7]

    Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei A. Efros. 2022. Test-time training with masked autoencoders. In NeurIPS 2022

    work page 2022

  8. [8]

    Alberto Garc \' a - Dur \' a n, Sebastijan Dumancic, and Mathias Niepert. 2018. Learning sequence encoders for temporal knowledge graph completion. In EMNLP 2018, pages 4816--4821. Association for Computational Linguistics

    work page 2018

Show all 43 references
  1. [9]

    Brubaker, and Pascal Poupart

    Rishab Goel, Seyed Mehran Kazemi, Marcus A. Brubaker, and Pascal Poupart. 2020. Diachronic embedding for temporal knowledge graph completion. In AAAI 2020, pages 3988--3995. AAAI Press

    2020

  2. [10]

    Goodfellow, Jean Pouget - Abadie, Mehdi Mirza, Bing Xu, David Warde - Farley, Sherjil Ozair, Aaron C

    Ian J. Goodfellow, Jean Pouget - Abadie, Mehdi Mirza, Bing Xu, David Warde - Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. 2014. https://arxiv.org/abs/1406.2661 Generative adversarial networks . CoRR, abs/1406.2661

    2014 arXiv

  3. [11]

    Zhen Han, Peng Chen, Yunpu Ma, and Volker Tresp. 2021 a . Explainable subgraph reasoning for forecasting on temporal knowledge graphs. In ICLR 2021. OpenReview.net

    2021

  4. [12]

    Zhen Han, Zifeng Ding, Yunpu Ma, Yujia Gu, and Volker Tresp. 2021 b . Learning neural ordinary equations for forecasting future links on temporal knowledge graphs. In EMNLP 2021, pages 8352--8364. Association for Computational Linguistics

    2021

  5. [13]

    Yongquan He, Peng Zhang, Luchen Liu, Qi Liang, Wenyuan Zhang, and Chuang Zhang. 2024. https://doi.org/10.48550/ARXIV.2402.12074 HIP network: Historical information passing network for extrapolation reasoning on temporal knowledge graph . CoRR, abs/2402.12074

    2024 doi arXiv

  6. [14]

    Yixin Ji, Kaixin Wu, Juntao Li, Wei Chen, Mingjie Zhong, Xu Jia, and Min Zhang. 2024. Retrieval and reasoning on kgs: Integrate knowledge graphs into large language models for complex question answering. In Findings of the Association for Computational Linguistics: EMNLP 2024,...

    2024

  7. [15]

    Tingsong Jiang, Tianyu Liu, Tao Ge, Lei Sha, Baobao Chang, Sujian Li, and Zhifang Sui. 2016. Towards time-aware knowledge graph completion. In COLING 2016, pages 1715--1724. ACL

    2016

  8. [16]

    Woojeong Jin, Meng Qu, Xisen Jin, and Xiang Ren. 2020. Recurrent event network: Autoregressive structure inferenceover temporal knowledge graphs. In EMNLP 2020, pages 6669--6683. Association for Computational Linguistics

    2020

  9. [17]

    Julien Leblay and Melisachew Wudage Chekol. 2018. Deriving validity time in knowledge graph. In WWW 2018, pages 1771--1776. ACM

    2018

  10. [18]

    Kalev Leetaru and Philip A Schrodt. 2013. Gdelt: Global data on events, location, and tone, 1979--2012. In ISA annual convention, volume 2, pages 1--49. Citeseer

    2013

  11. [19]

    Zixuan Li, Xiaolong Jin, Wei Li, Saiping Guan, Jiafeng Guo, Huawei Shen, Yuanzhuo Wang, and Xueqi Cheng. 2021. Temporal knowledge graph reasoning based on evolutional representation learning. In SIGIR '21, pages 408--417. ACM

    2021

  12. [20]

    Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot - Gurlet, Taylor Mordan, and Alexandre Alahi. 2021. TTT++: when does self-supervised test-time training fail or thrive? In NeurIPS 2021, pages 21808--21820

    2021

  13. [21]

    Yushan Liu, Yunpu Ma, Marcel Hildebrandt, Mitchell Joblin, and Volker Tresp. 2022. Tlogic: Temporal logical rules for explainable link forecasting on temporal knowledge graphs. In AAAI 2022, pages 4120--4127. AAAI Press

    2022

  14. [22]

    Suchanek

    Farzaneh Mahdisoltani, Joanna Biega, and Fabian M. Suchanek. 2015. YAGO3: A knowledge base from multilingual wikipedias. In Seventh Biennial Conference on Innovative Data Systems Research, CIDR 2015 . www.cidrdb.org

    2015

  15. [23]

    Lingyuan Meng, Ke Liang, Hao Yu, Yue Liu, Sihang Zhou, Meng Liu, and Xinwang Liu. 2024. Fedean: Entity-aware adversarial negative sampling for federated knowledge graph reasoning. IEEE Trans. Knowl. Data Eng. , 36(12):8206--8219

    2024

  16. [24]

    Namyong Park, Fuchen Liu, Purvanshi Mehta, Dana Cristofor, Christos Faloutsos, and Yuxiao Dong. 2022. Evokg: Jointly modeling event time and network structure for reasoning over temporal knowledge graphs. In WSDM '22, pages 794--803. ACM

    2022

  17. [25]

    Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, and 2 others

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas K \" o pf, Edward Z. Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Stei...

    2019

  18. [26]

    Krinski, and Eduardo Todt

    Daniel Vitor Ruiz, Bruno A. Krinski, and Eduardo Todt. 2019. ANDA: A novel data augmentation technique applied to salient object detection. In ICAR 2019, pages 487--492. IEEE

    2019

  19. [27]

    Shahriar Shayesteh and Diana Inkpen. 2022. Generative adversarial learning with negative data augmentation for semi-supervised text classification. In FLAIRS 2022

    2022

  20. [28]

    Yuehang Si, Xingchen Hu, Qing Cheng, Xinwang Liu, Shixuan Liu, and Jincai Huang. 2025. Coherence mode: Characterizing local graph structural information for temporal knowledge graph. Inf. Sci., 686:121357

    2025

  21. [29]

    Haohai Sun, Jialun Zhong, Yunpu Ma, Zhen Han, and Kun He. 2021. Timetraveler: Reinforcement learning for temporal knowledge graph forecasting. In EMNLP 2021, pages 8306--8319. Association for Computational Linguistics

    2021

  22. [30]

    Efros, and Moritz Hardt

    Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei A. Efros, and Moritz Hardt. 2019 a . https://arxiv.org/abs/1909.13231 Test-time training for out-of-distribution generalization . CoRR, abs/1909.13231

    2019 arXiv

  23. [31]

    Efros, and Moritz Hardt

    Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei A. Efros, and Moritz Hardt. 2020. Test-time training with self-supervision for generalization under distribution shifts. In ICML 2020, volume 119, pages 9229--9248. PMLR

    2020

  24. [32]

    Zhiqing Sun, Zhi - Hong Deng, Jian - Yun Nie, and Jian Tang. 2019 b . Rotate: Knowledge graph embedding by relational rotation in complex space. In ICLR 2019. OpenReview.net

    2019

  25. [33]

    Rakshit Trivedi, Hanjun Dai, Yichen Wang, and Le Song. 2017. Know-evolve: Deep temporal reasoning for dynamic knowledge graphs. In ICML 2017, volume 70, pages 3462--3471. PMLR

    2017

  26. [34]

    Th \' e o Trouillon, Johannes Welbl, Sebastian Riedel, \' E ric Gaussier, and Guillaume Bouchard. 2016. Complex embeddings for simple link prediction. In ICML 2016, volume 48, pages 2071--2080. JMLR.org

    2016

  27. [35]

    Zhen Wang, Jianwen Zhang, Jianlin Feng, and Zheng Chen. 2014. Knowledge graph embedding by translating on hyperplanes. In Proceedings of the AAAI conference on artificial intelligence, volume 28

    2014

  28. [36]

    Yi Xu, Junjie Ou, Hui Xu, and Luoyi Fu. 2023. Temporal knowledge graph reasoning with historical contrastive learning. In AAAI 2023, pages 4765--4773. AAAI Press

    2023

  29. [37]

    Bishan Yang, Wen - tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. 2015. Embedding entities and relations for learning and inference in knowledge bases. In ICLR 2015

    2015

  30. [38]

    Xihong Yang, Yiqi Wang, Jin Chen, Wenqi Fan, Xiangyu Zhao, En Zhu, Xinwang Liu, and Defu Lian. 2024. https://arxiv.org/abs/2407.15620 Dual test-time training for out-of-distribution recommender system . CoRR, abs/2407.15620

    2024 arXiv

  31. [39]

    Lupeng Yue, Yongjian Ren, Yan Zeng, Jilin Zhang, Kaisheng Zeng, Jian Wan, and Mingyao Zhou. 2024. Complex expressional characterizations learning based on block decomposition for temporal knowledge graph completion. Knowl. Based Syst., 290:111591

    2024

  32. [40]

    Jiaxin Zhang, Yiqi Wang, Xihong Yang, Siwei Wang, Yu Feng, Yu Shi, Ruichao Ren, En Zhu, and Xinwang Liu. 2024. Test-time training on graphs with large language models (llms). In ACM MM 2024, pages 2089--2098. ACM

    2024

  33. [41]

    Cunchao Zhu, Muhao Chen, Changjun Fan, Guangquan Cheng, and Yan Zhang. 2021. Learning from history: Modeling temporal knowledge graphs with sequential copy-generation networks. In AAAI 2021, pages 4732--4740. AAAI Press

    2021

  34. [42]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  35. [43]

    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 6, 2026 · model on record in the stance chip above.