Pith. sign in

REVIEW 4 major objections 5 minor 28 references

ETT-CKGE: Efficient Task-driven Tokens for Continual Knowledge Graph Embedding

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Learnable task-driven tokens replace hand-crafted importance scores in continual knowledge graph embedding, cutting training time by 50-96% while matching or exceeding link prediction accuracy.

desk verdict Genuinely new efficiency trick for continual KG embedding with real training-time gains, but the accuracy advantage is tiny and the headline 50–96% speedup is dataset-dependent; deserves peer review with revisions. read the letter →

arxiv 2506.08158 v1 pith:BHVK5RMX submitted 2025-06-09 cs.CL

Lijing Zhu , Qizhen Lan , Qing Tian , Wenbo Sun , Li Yang , Lu Xia , Yixin Xie , Xi Xiao
show 3 more authors
Tiehang Duan Cui Tao Shuteng Niu
This is my paper
classification cs.CL
keywords continualknowledgegraphembeddingtask-driventokensdistillationlinkpredictioncatastrophicforgettingtraversal-freeregularization-basedlearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that continual knowledge graph embedding does not need manually designed node and relation importance scores, which are both suboptimal and computationally costly. The authors propose replacing them with a small set of learnable tokens that produce a soft importance mask through one matrix multiplication, trained directly by the link prediction loss. On six benchmark datasets, this approach reportedly achieves the same or better mean reciprocal rank than prior state-of-the-art methods while cutting training time by 50-96% relative to the second-best performer and using less memory. If correct, ETT-CKGE offers a scalable, traversal-free way to update knowledge graph embeddings over time, making continual learning practical for larger graphs.

What carries the argument

The central object is the task-driven token mask: a set of $T$ learnable vectors $Z$ that, when multiplied by an embedding matrix, produce a soft importance mask $M = \sigma(Z E^\top)$. The mask is applied as a token-wise weighted sum to create a masked embedding, and an aligned joint mask $M_{i-1} \odot M_i$ drives the distillation loss between snapshots. This machinery replaces heuristic importance scoring and graph traversal with a single matrix multiplication, which is what yields the reported efficiency gains.

What would settle it

One concrete experiment would be to take a dataset such as RELATION, where relations grow sharply between snapshots, and compare ETT-CKGE with tokens frozen from snapshot 0 against a variant that retrains the tokens at every snapshot. If the retrained-token variant achieves substantially higher MRR, the claim that frozen tokens remain valid cross-snapshot guidance would be falsified.

Watch

Extended reading notes

Core claim

ETT-CKGE claims that the importance of entities and relations for continual knowledge transfer can be learned end-to-end from the task objective, rather than approximated by heuristics. The method introduces learnable tokens $Z \in \mathbb{R}^{T\times D}$ that interact with the old snapshot embeddings $E_{i-1}$ through a sigmoid-activated inner product to form a soft mask $M_{i-1} = \sigma(Z E_{i-1}^\top)$. In Stage I the tokens are trained, together with a diversity loss based on the Dice coefficient, to capture task-relevant and distinct graph substructures. In Stage II the tokens and old embeddings are frozen, and a joint mask $M = M_{i-1} \odot M_i$ is applied to both old and new embeddings to weight an L2 distillation loss, aligning the new embedding with the old one only on components both snapshots consider important. Because importance estimation reduces to simple matrix operations, the method avoids graph traversal and iterative scoring entirely, and the paper reports consistent 50-96% training-time reductions against the second-best baseline while maintaining or improving MRR and lowering peak memory use.

Load-bearing premise

The load-bearing premise is that tokens learned on the old snapshot remain valid guidance for the new snapshot once frozen, so if the set of important entities and relations shifts, the aligned mask could suppress newly important components and the distillation would preserve the wrong parts of the old embedding.

Editorial extensions

If this is right

  • If correct, continual knowledge graph embedding no longer needs per-node and per-relation importance scoring, so training time can scale with the number of embeddings rather than with costly graph traversal.
  • The frozen token masks serve as reusable guidance across snapshots, meaning the same tokens can steer distillation without being retrained, simplifying continual learning pipelines.
  • The 50-96% training-time reduction over the second-best method indicates that prior importance computation, not the embedding update itself, is the dominant computational bottleneck in CKGE.
  • The method updates only about 2,000 parameters per snapshot, suggesting that continual knowledge graph embedding can be achieved with a fixed, lightweight parameter budget.

Reading between the lines

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

  • An implicit consequence is that the learned soft masks could double as interpretable importance scores, letting practitioners see which entities and relations the model considers critical at each snapshot; the paper does not analyze this directly.
  • The paper does not separately validate the frozen-token transfer assumption: a natural test is to compare MRR when tokens are frozen versus retrained at each snapshot on a relation-heavy dataset like RELATION, where the set of important relations changes quickly.
  • Since the diversity loss encourages token specialization, the choice of token count $T$ likely interacts with the number of distinct graph substructures; the paper only reports tuning $T$ in a small range and does not study that interaction.
  • The token-masking idea could extend beyond knowledge graphs to other evolving graph tasks such as node classification, where the same aligned-mask distillation might replace heuristic importance weighting; this is an editorial extrapolation, not a claim in the paper.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes ETT-CKGE, a continual knowledge graph embedding method in which a set of learnable task-driven tokens interact with old-snapshot embeddings to produce importance masks. Stage I (Section 4.1) trains the tokens with a TransE task loss plus a Dice-based diversity loss; Stage II (Section 4.2) freezes the tokens and old embeddings, trains the new-snapshot embeddings with the task loss and a token-masked distillation loss, and uses an intersection mask M = M_{i-1} ⊙ M_i to align old and new embeddings. The authors evaluate on six CKGE benchmarks and report MRR/Hits@k, training time, memory, and ablations, claiming a 50–96% training-time reduction at comparable or superior MRR compared with state-of-the-art baselines.

Significance. If the empirical claims hold, the paper makes a useful contribution: it replaces heuristic node/relation importance scores with masks learned directly from the task objective, reduces the graph-traversal overhead of prior regularization-based CKGE methods to matrix operations, and provides public code and an ablation study isolating the distillation loss. The main weaknesses are that the central mechanism (mask transfer across snapshots) is not directly validated, the small MRR differences are reported without uncertainty, and the hyperparameter selection protocol is underspecified. These issues are fixable with additional experiments and reporting, so the result is promising but not yet fully substantiated.

major comments (4)
  1. [4.2, Eq. (7)] In Eq. (7), the aligned mask M = M_{i-1} ⊙ M_i depends on the trainable new embeddings through M_i = σ(Z E_i^T). As E_i is optimized in Stage II, components that become newly important in the current snapshot have small M_i entries and are therefore down-weighted in L_distill; old knowledge stored in those components is then protected only by L_trans, which is exactly the forgetting channel. Figure 4 reports only final MRR and Table 4 ablates L_distill as a whole, so neither isolates whether the intersection mask transfers correctly. Please add a direct diagnostic, for example comparing L_distill variants with M_{i-1} only, with a union mask, or with no mask, and reporting old-snapshot MRR after each stage; alternatively, report a per-component correlation between the learned mask and measured forgetting. This is load-bearing for the reusable-guidance claim and for the 'comparable/superior MRR' half of the headline result.
  2. [5.5, Table 3] The headline 50–96% training-time reduction at 'comparable or superior MRR' is not accompanied by uncertainty quantification. Several accuracy differences are very small (e.g., RELATION MRR 0.200 vs IncDE 0.199; FACT 0.217 vs IncDE 0.216; HYBRID 0.224 vs IncDE 0.223), and on WN-CKGE the ETT-CKGE MRR of 0.153 is below FastKGE's 0.160, with the text describing this only as competitive. Since Section 5.3 states that every result is the average of five runs, report standard deviations or confidence intervals, and specify per dataset which model is the 'second-best performing model' used to compute the 50–96% range.
  3. [5.3] The hyperparameter selection protocol for ETT-CKGE is not described separately from the baselines. The text gives ranges for α (1,000 to 100,000), λ ([0,1]), T (integer in (0,10]), batch size, and learning rate, but does not state how these were chosen per dataset, whether a validation split was used for model selection, or whether baseline hyperparameters were tuned with the same protocol. Because the efficiency comparison depends on the reported training times, please provide per-dataset settings and the selection procedure, or state that a fixed configuration was used.
  4. [5.6, Fig. 4] The paper's motivation is mitigating catastrophic forgetting, but Fig. 4 reports only final MRR per snapshot and the text comments on heatmap patterns without a quantitative forgetting metric. Report a standard measure such as average forgetting or backward transfer (e.g., the difference between peak performance and final performance on earlier snapshots) to support the claim of effective knowledge preservation, especially on the RELATION dataset where snapshot-to-snapshot variation is visible.
minor comments (5)
  1. [Section 4] The section title 'Efficient Task-diven Tokens' should read 'Task-driven'; the typo appears in the heading.
  2. [Section 5.8] The abbreviation 'Ldist' is used in the paragraph on the diversity loss ('not as crucial as Ldist') where L_distill is meant.
  3. [Section 5.7] The bullet for Updated Parameters is missing its bullet marker in the text.
  4. [Table 4] The abbreviation SIT is defined only in the body text; spell out 'Stage I Token training' in the caption for readability.
  5. [Section 5.7] The section says the analysis covers 'snapshot 2 onward,' but the first sentence says Figure 5 provides validation on the RELATION dataset; clarify whether the plotted metrics are cumulative from snapshot 2 or from snapshot 1.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: ETT-CKGE's token masks are trained on the task loss and evaluated on held-out link prediction, and the only self-citation is a baseline comparison that is not load-bearing.

full rationale

The paper's derivation chain is empirical rather than definitional. Stage-I tokens Z are learned by optimizing L_token = L_trans + λL_div on the previous snapshot's embeddings, producing importance masks via Eq. (1). These masks are then frozen and reused in Eq. (6)-(7) to weight distillation, while the overall Stage-II objective in Eq. (8) is still optimized against the link-prediction task loss. Final MRR and Hits@k are measured on held-out test facts, so no fitted parameter is relabeled as an independent prediction. The efficiency claims (training time, memory, updated parameters) are directly measured against baselines, not derived from the method's own assumptions. The one overlapping-author citation is FMR [29], which appears only as a compared baseline in Section 5.2 and Table 3 and as a related-work mention; it is not used to justify ETT-CKGE's design, and the paper does not import any uniqueness theorem or ansatz from prior work. The identified risk that frozen tokens may become stale when embeddings shift (Eqs. 6-7) is a robustness concern about mask transfer, not a circularity, because the final evaluation still tests the complete model on unseen data. Therefore no specific equation or claim reduces by construction to its own input, and the paper is appropriately scored as essentially non-circular, with only a negligible self-citation that does not affect the central result.

Assumptions & free parameters 7 free parameters · 5 assumptions · 1 invented entities

The method's central claim depends mainly on design assumptions rather than physical constants. The free parameters are per-dataset hyperparameters and the token set is a learned latent construct. The most consequential premise is mask transferability across snapshots, which is asserted and tested only indirectly through final link prediction metrics.

free parameters (7)
  • alpha (distillation weight) = 1,000 to 100,000 per dataset
    Controls L_distill in Eq. (8); selected per dataset from a wide range.
  • lambda (diversity weight) = [0, 1] per dataset
    Balances L_div in Eq. (5); no fixed selection rule given.
  • token count T = integer in (0, 10] per dataset
    Number of learnable tokens; sets the capacity of the importance masks.
  • learning rate = {0.01, 0.001, 0.0001, 0.00001}
    Chosen from a grid during tuning.
  • batch size = {1024, 2048, 3072}
    Chosen from a grid during tuning.
  • margin gamma = 9
    Margin in the TransE loss, Eq. (4), fixed for all experiments.
  • embedding dimension D = 200
    Dimensionality of entity, relation, and token embeddings; set once for all experiments.
assumptions (5)
  • domain assumption TransE scoring is an adequate task loss for continual link prediction in these benchmarks.
    Used in Eq. (4) for both token learning and final training; no comparison with other scorers is made within ETT-CKGE.
  • domain assumption Frozen token masks trained on the previous snapshot remain useful for the current snapshot.
    Stage II freezes Z and old embeddings; the paper asserts transferability in Section 4.2 but does not separately measure it.
  • domain assumption Element-wise product of old and new masks identifies the components worth distilling.
    M = M_{i-1} elementwise M_i in Section 4.2 is a design choice, not derived from an optimality argument.
  • domain assumption Dice-based diversity loss makes tokens specialize on different substructures.
    Eq. (3) penalizes mask overlap; no analysis shows that the learned masks are actually diverse.
  • ad hoc to paper New entities and relations need no distillation, only task loss.
    Section 4.2 excludes new entities and relations from distillation to avoid interference; this is necessary for the method but not independently justified.
invented entities (1)
  • Task-driven token set Z (T x D learnable vectors)
    purpose: Learned soft importance masks for entities and relations; replaces handcrafted importance metrics and graph traversal.
    Tokens are model parameters trained on benchmark losses; there is no external falsifiable handle or prior theoretical grounding, so their transferability rests on the paper's experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ETT-CKGE: Efficient Task-driven Tokens for Continual Knowledge Graph Embedding." pith.science (2026). https://pith.science/paper/BHVK5RMX

@misc{pith2026250608158,
  author       = {Pith},
  title        = {Pith review of: ETT-CKGE: Efficient Task-driven Tokens for Continual Knowledge Graph Embedding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BHVK5RMX}},
  note         = {Machine review of arXiv:2506.08158}
}
read the original abstract

Continual Knowledge Graph Embedding (CKGE) seeks to integrate new knowledge while preserving past information. However, existing methods struggle with efficiency and scalability due to two key limitations: (1) suboptimal knowledge preservation between snapshots caused by manually designed node/relation importance scores that ignore graph dependencies relevant to the downstream task, and (2) computationally expensive graph traversal for node/relation importance calculation, leading to slow training and high memory overhead. To address these limitations, we introduce ETT-CKGE (Efficient, Task-driven, Tokens for Continual Knowledge Graph Embedding), a novel task-guided CKGE method that leverages efficient task-driven tokens for efficient and effective knowledge transfer between snapshots. Our method introduces a set of learnable tokens that directly capture task-relevant signals, eliminating the need for explicit node scoring or traversal. These tokens serve as consistent and reusable guidance across snapshots, enabling efficient token-masked embedding alignment between snapshots. Importantly, knowledge transfer is achieved through simple matrix operations, significantly reducing training time and memory usage. Extensive experiments across six benchmark datasets demonstrate that ETT-CKGE consistently achieves superior or competitive predictive performance, while substantially improving training efficiency and scalability compared to state-of-the-art CKGE methods. The code is available at: https://github.com/lijingzhu1/ETT-CKGE/tree/main

Figures

Figures reproduced from arXiv: 2506.08158 by the authors.

Figure 1
Figure 1. (Best view in color) Comparison of performance(MRR), training time(S), and memory usage(MB) across CKGE methods on the RELATION data set. Our method achieves the best balance, delivering high accuracy with significantly reduced training time and memory consumption. The color scale indicates memory usage, with darker colors representing lower memory overhead. sal and depend on handcrafted metrics to guide learning. I… view at source ↗
Figure 2
Figure 2. An overview of the ETT-CKGE framework. Stage I focuses on token pre [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Entity embedding visualization on the ENTITY dataset. The Silhouette Score, [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: MRR Changes across six datasets during sequential learning. Each heatmap displays the MRR achieved by our model on each dataset over snapshots S1 to S5. Model stages are denoted as M i, where i represents the snapshot number. Warmer colors indicate higher MRR, while co…
Figure 5
Figure 5. Figure 5: Model Scalability Analysis Over Time the sharp cost spikes observed in other models. Such smooth transitions sug￾gest that the task-driven token mechanism effectively captures and reuses informative components without requiring heavy computational overhead. While train…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 13 canonical work pages

  1. [1]

    Advances in neural information processing systems26(2013)

    Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J., Yakhnenko, O.: Translating embeddings for modeling multi-relational data. Advances in neural information processing systems26(2013)

    2013

  2. [2]

    Advances in neural information processing systems33, 1877–1901 (2020)

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Nee- lakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems33, 1877–1901 (2020)

    2020

  3. [3]

    Journal of Machine Learning Research24(240), 1–113 (2023)

    Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling lan- guage modeling with pathways. Journal of Machine Learning Research24(240), 1–113 (2023)

    work page 2023

  4. [4]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Cui,Y.,Wang,Y.,Sun,Z.,Liu,W.,Jiang,Y.,Han,K.,Hu,W.:Lifelongembedding learning and transfer for growing knowledge graphs. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 4217–4224 (2023)

    work page 2023

  5. [6]

    In: International conference on machine learning

    Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904–4916. PMLR (2021)

    2021

  6. [7]

    Neural Networks p

    Ju, W., Fang, Z., Gu, Y., Liu, Z., Long, Q., Qiao, Z., Qin, Y., Shen, J., Sun, F., Xiao, Z., et al.: A comprehensive survey on deep graph representation learning. Neural Networks p. 106207 (2024)

    work page 2024

  7. [8]

    Advances in neural information processing systems31(2018)

    Kazemi,S.M.,Poole,D.:Simpleembeddingforlinkpredictioninknowledgegraphs. Advances in neural information processing systems31(2018)

    work page 2018

  8. [9]

    Proceedings of the national academy of sciences114(13), 3521–3526 (2017)

    Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences114(13), 3521–3526 (2017)

    2017

Show all 28 references
  1. [10]

    Journal of Biomedical Informatics p

    Li, Y., Viswaroopan, D., He, W., Li, J., Zuo, X., Xu, H., Tao, C.: Improving entity recognition using ensembles of deep learning and fine-tuned large language models: A case study on adverse event extraction from vaers and social media. Journal of Biomedical Informatics p. 104...

    2025

  2. [11]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Liu, J., Ke, W., Wang, P., Shang, Z., Gao, J., Li, G., Ji, K., Liu, Y.: Towards continual knowledge graph embedding via incremental distillation. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 8759–8768 (2024)

    2024

  3. [12]

    arXiv preprint arXiv:2407.05705 (2024)

    Liu, J., Ke, W., Wang, P., Wang, J., Gao, J., Shang, Z., Li, G., Xu, Z., Ji, K., Li, Y.: Fast and continual knowledge graph embedding via incremental lora. arXiv preprint arXiv:2407.05705 (2024)

    2024 arXiv

  4. [13]

    In: 2018 24th International Conference on Pattern Recognition (ICPR)

    Liu, X., Masana, M., Herranz, L., Van de Weijer, J., Lopez, A.M., Bagdanov, A.D.: Rotate your networks: Better weight consolidation and less catastrophic forgetting. In: 2018 24th International Conference on Pattern Recognition (ICPR). pp. 2262–

    2018

  5. [14]

    In: Proceedings of the ACM web conference 2023

    Liu, Z., Yu, X., Fang, Y., Zhang, X.: Graphprompt: Unifying pre-training and downstream tasks for graph neural networks. In: Proceedings of the ACM web conference 2023. pp. 417–428 (2023) 16 L. Zhu at al

    2023

  6. [15]

    In: Conference on robot learning

    Lomonaco, V., Maltoni, D.: Core50: a new dataset and benchmark for continuous object recognition. In: Conference on robot learning. pp. 17–26. PMLR (2017)

    2017

  7. [16]

    Advances in neural information processing systems30(2017)

    Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. Advances in neural information processing systems30(2017)

    2017

  8. [17]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021)

    2021

  9. [18]

    arXiv preprint arXiv:1606.04671 (2016)

    Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016)

    2016 arXiv

  10. [19]

    arXiv preprint arXiv:2302.13971 (2023)

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    2023 arXiv

  11. [20]

    In: International conference on machine learning

    Trouillon, T., Welbl, J., Riedel, S., Gaussier, É., Bouchard, G.: Complex embed- dings for simple link prediction. In: International conference on machine learning. pp. 2071–2080. PMLR (2016)

    2016

  12. [21]

    arXiv preprint arXiv:1903.02588 (2019)

    Wang, H., Xiong, W., Yu, M., Guo, X., Chang, S., Wang, W.Y.: Sentence embed- ding alignment for lifelong relation extraction. arXiv preprint arXiv:1903.02588 (2019)

    2019 arXiv

  13. [22]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Wang, P., Han, J., Li, C., Pan, R.: Logic attention based neighborhood aggregation for inductive knowledge graph embedding. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 33, pp. 7152–7159 (2019)

    2019

  14. [23]

    Wang, Q., Mao, Z., Wang, B., Guo, L.: Knowledge graph embedding: A survey of approachesandapplications.IEEEtransactionsonknowledgeanddataengineering 29(12), 2724–2743 (2017)

    2017

  15. [24]

    arXiv preprint arXiv:2406.17697 (2024)

    Xiao,X.,Wang,W.,Xie,J.,Zhu,L.,Chen,G.,Li,Z.,Wang,T.,Xu,M.:Hgtdp-dta: Hybrid graph-transformer with dynamic prompt for drug-target binding affinity prediction. arXiv preprint arXiv:2406.17697 (2024)

    2024 arXiv

  16. [25]

    arXiv preprint arXiv:2203.15827 (2022)

    Yasunaga, M., Leskovec, J., Liang, P.: Linkbert: Pretraining language models with document links. arXiv preprint arXiv:2203.15827 (2022)

    2022 arXiv

  17. [26]

    arXiv preprint arXiv:1708.01547 (2017)

    Yoon, J., Yang, E., Lee, J., Hwang, S.J.: Lifelong learning with dynamically ex- pandable networks. arXiv preprint arXiv:1708.01547 (2017)

    2017 arXiv

  18. [27]

    arXiv preprint arXiv:2402.11565 (2024)

    Zhang, X., Song, D., Tao, D.: Continual learning on graphs: Challenges, solutions, and opportunities. arXiv preprint arXiv:2402.11565 (2024)

    2024 arXiv

  19. [28]

    In: Proceedings of the AAAI Conference on Artificial In- telligence

    Zhou, F., Cao, C.: Overcoming catastrophic forgetting in graph neural networks with experience replay. In: Proceedings of the AAAI Conference on Artificial In- telligence. vol. 35, pp. 4714–4722 (2021)

    2021

  20. [29]

    In: 2024 IEEE International Conference on Big Data (BigData)

    Zhu, L., Jeon, D.H., Sun, W., Yang, L., Xie, Y., Niu, S.: Flexible memory ro- tation (fmr): Rotated representation with dynamic regularization to overcome catastrophic forgetting in continual knowledge graph learning. In: 2024 IEEE International Conference on Big Data (BigData...

    2024

Pith tools

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