arxiv: 2605.10529 · v1 · submitted 2026-05-11 · 💻 cs.AI · cs.LG

Recognition: no theorem link

PrimeKG-CL: A Continual Graph Learning Benchmark on Evolving Biomedical Knowledge Graphs

Yousef A. Radwan , Yao Li , Qing Qing , Ziqi Xu , Xingtong Yu , Jiaxing Huang , Renqiang Luo , Xikun Zhang

Authors on Pith no claims yet

Pith reviewed 2026-05-12 04:32 UTC · model grok-4.3

classification 💻 cs.AI cs.LG

keywords continual graph learningbiomedical knowledge graphsknowledge graph embeddingsbenchmarktemporal snapshotsentity classificationrelationship prediction

0 comments

The pith

Real updates to biomedical knowledge graphs show that decoder and continual learning strategy must be matched to avoid performance loss.

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

The paper builds PrimeKG-CL from two actual releases of a large biomedical knowledge graph to test continual learning methods under realistic conditions. Existing work relies on artificial splits of static graphs, but this benchmark uses millions of added and removed edges from nine source databases along with task stratifications for persistent, new, and deprecated knowledge. Evaluations across six strategies and four decoders on relationship prediction, entity classification, and question answering tasks reveal that no strategy works best for all decoders and that only the DistMult decoder distinguishes between keeping valid facts and forgetting outdated ones. Multimodal features help on entity tasks, but some methods do not scale to the full size. This matters for keeping biomedical AI tools current as ontologies change.

Core claim

The core discovery is that decoder choice and continual learning strategy interact strongly on this benchmark: mismatched combinations degrade performance, and only DistMult shows clear separation between persistent and deprecated knowledge, while standard metrics conflate the two; multimodal features improve entity-level performance by up to 60%.

What carries the argument

PrimeKG-CL benchmark with its two temporal snapshots from June 2021 and July 2023, 10 entity-type-grouped tasks, and per-task persistent/added/removed test stratification.

If this is right

No single continual learning strategy performs best across all knowledge graph embedding decoders.
Mismatched decoder-strategy pairs can significantly degrade performance on biomedical tasks.
Only the DistMult decoder exhibits separation between persistent and deprecated knowledge.
Multimodal node features improve entity classification and related tasks by up to 60%.
Certain continual knowledge graph embedding frameworks fail to scale to the 5.67 million triple base task.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This suggests that practical systems for drug repurposing or clinical support using knowledge graphs will require careful pairing of embedding methods with update strategies.
The findings point to a need for new evaluation metrics that explicitly measure retention versus forgetting in evolving graphs.
The benchmark could be extended with more frequent snapshots to study longer-term continual learning dynamics.

Load-bearing premise

The two chosen snapshots and the entity-type-grouped task definitions sufficiently represent the asynchronous, structured evolution that real biomedical KGs undergo in practice.

What would settle it

If evaluations on this benchmark showed similar performance for all decoder and strategy combinations on the persistent versus removed triples, or if DistMult failed to separate them, the claims about interactions and metric issues would be falsified.

Figures

Figures reproduced from arXiv: 2605.10529 by Jiaxing Huang, Qing Qing, Renqiang Luo, Xikun Zhang, Xingtong Yu, Yao Li, Yousef A. Radwan, Ziqi Xu.

**Figure 1.** Figure 1: PrimeKG-CL knowledge evolution from t0 (June 2021) to t1 (July 2023): 5.83M new edges, 889K deprecated, 7.21M persistent. Node size encodes per-entity-type frequency at each snapshot. In contrast to prior CGL benchmarks, PrimeKG-CL provides the first evaluation framework on a real-world biomedical knowledge graph with genuine temporal evolution and multimodal node features, encompassing three evaluation tr… view at source ↗

**Figure 2.** Figure 2: The 10-task continual-learning sequence on PrimeKG-CL (top) and per-task memory [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Decoder × CL interaction (AP relative to Naive, per column): EWC transforms ComplEx (+167%); SI/Distillation drop RotatE by 37%/57%; replay drops ComplEx by 38%. Per-task dynamics and the difficulty of real biomedical KGs. Per-task peak MRR reaches ≥ 0.25 on several tasks (e.g., CMKL: Disease-r1 0.32, Gene/Protein-r2 0.31, Phenotype-r3 0.25), confirming meaningful patterns are learned even when AP is low. … view at source ↗

**Figure 4.** Figure 4: Per-task peak MRR for CMKL, Naive, and Joint; grey line [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Per-task forgetting (peak − final MRR). CMKL concentrates forgetting on Disease-r1 (56%, 0.227); Naive/EWC concentrate it on Base-t0 (70%). 4.4 Biomedical Knowledge Graph Question Answering (KGQA) Results We position the KGQA track as a stress test rather than a finished evaluation. All three configurations of Qwen2.5-7B-Instruct sit at or below token F1 = 0.015, with full RAG (≤ 0.015) only marginally abo… view at source ↗

read the original abstract

Biomedical knowledge graphs underwrite drug repurposing and clinical decision support, yet the upstream ontologies they depend on update on independent cycles that add millions of edges and deprecate hundreds of thousands more between releases. Yet existing continual graph learning has been studied almost exclusively on synthetic random splits of static, generic KGs, a regime that cannot reproduce the asynchronous, structured evolution real biomedical KGs undergo. To this end, we introduce PrimeKG-CL, a CGL benchmark built from nine authoritative biomedical databases (129K+ nodes, 8.1M+ edges, 10 node types, 30 relation types) with two genuine temporal snapshots (June 2021, July 2023; 5.83M edges added, 889K removed, 7.21M persistent), 10 entity-type-grouped tasks, multimodal node features, and a per-task persistent/added/removed test stratification. On three tasks (biomedical relationship prediction, entity classification, KGQA), we evaluate six CL strategies across four KGE decoders, plus LKGE, an LLM-RAG agent, and CMKL. We find that decoder choice and continual learning strategy interact strongly: no single strategy performs best across all decoders, and mismatched combinations can significantly degrade performance. Moreover, only DistMult exhibits a clear separation between persistent and deprecated knowledge, indicating that standard metrics conflate retention of still-valid facts with failure to forget outdated ones; this effect is absent under RotatE. In addition, multimodal features improve entity-level tasks by up to 60%, and a recent CKGE framework (IncDE) failed to scale to our 5.67M-triple base task across five attempts up to 350GB RAM. Data, pipeline, baselines, and the stratified split are released openly. Dataset:huggingface.co/datasets/yradwan147/PrimeKGCL|Code:github.com/yradwan147/primekg-cl-neurips2026

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Real temporal biomedical KG benchmark shows decoder-CL strategy interactions but is limited to one transition between two snapshots.

read the letter

The main things here are that PrimeKG-CL is the first benchmark using actual temporal snapshots from real biomedical databases instead of random synthetic splits on static graphs, and the experiments turn up clear interactions between embedding decoders and continual learning strategies that synthetic setups have missed. They built it from nine sources with June 2021 and July 2023 snapshots, millions of added and removed edges, ten entity-type tasks, and test sets split into persistent, added, and removed facts. They ran six CL strategies across four KGE decoders on relationship prediction, entity classification, and KGQA, plus a couple of extras. The results show no single strategy works best across decoders, mismatched pairs can hurt, and only DistMult separates still-valid facts from failure to drop outdated ones; multimodal features help on entity tasks and one prior method did not scale. The open data, code, and stratified splits are a real plus for anyone who wants to test new methods on this domain. The two-snapshot setup is the clearest soft spot. Continual learning usually implies repeated updates over time, but here it reduces to a single transition, so the interaction effects and the DistMult-specific behavior might not hold when changes accumulate across multiple releases or when update magnitudes vary. The entity-type task grouping is reasonable but may not capture all the asynchronous ontology shifts that happen in practice. The abstract reports concrete numbers, yet without the full tables and split-construction details it is hard to judge how robust the separation claims are. This is for researchers working on continual graph learning who care about applied, evolving domains rather than generic benchmarks. It has enough new construction and empirical observations to deserve a serious referee even if the single-transition design needs more discussion in review.

Referee Report

2 major / 2 minor

Summary. The paper introduces PrimeKG-CL, a continual graph learning benchmark constructed from the PrimeKG biomedical knowledge graph using two temporal snapshots (June 2021 and July 2023). It features 129K+ nodes, 8.1M+ edges, 10 node types, and 30 relation types, with 10 entity-type-grouped tasks, multimodal node features, and stratified test sets for persistent, added, and removed knowledge. Experiments on biomedical relationship prediction, entity classification, and KGQA evaluate six CL strategies across four KGE decoders, plus additional baselines like LKGE and CMKL. The key results are strong interactions between decoder choice and CL strategy, with mismatched combinations degrading performance, and only DistMult showing clear separation between persistent and deprecated knowledge. Multimodal features improve performance by up to 60%, and some frameworks fail to scale. The dataset, code, and splits are released openly.

Significance. If these results hold, the benchmark fills an important gap by providing a realistic setting for CGL on evolving biomedical KGs, unlike prior work on synthetic splits. The findings on decoder-CL interactions and the inadequacy of standard metrics for distinguishing retention from forgetting are valuable for guiding future method development. The open data and code release is a notable strength that promotes reproducibility and community use. This could have implications for applications in drug repurposing and clinical decision support where KGs evolve over time.

major comments (2)

[§3 (Benchmark Construction)] §3 (Benchmark Construction): The benchmark relies on only two snapshots, corresponding to a single transition from June 2021 to July 2023. This setup does not capture cumulative effects, repeated updates, or varying change magnitudes typical in real biomedical KGs with independent ontology cycles. Consequently, the observed strong decoder-CL strategy interactions and the DistMult-specific separation between persistent and deprecated knowledge may be artifacts of this one-step regime rather than general properties.
[§4 (Experiments)] §4 (Experiments): Full details on hyperparameter choices, statistical significance testing across runs, and the precise construction of the 10 entity-type-grouped tasks with per-task persistent/added/removed stratification are not elaborated, which is load-bearing for assessing whether the reported decoder-strategy interactions are robust.

minor comments (2)

[Abstract] Abstract: The dataset and code links are concatenated without spaces or separators ('Dataset:huggingface.co/datasets/yradwan147/PrimeKGCL|Code:github.com/yradwan147/primekg-cl-neurips2026').
[Throughout] Throughout: Terminology for 'persistent', 'added', and 'removed' knowledge should be used consistently in text, tables, and figure captions to avoid ambiguity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed and constructive review of our manuscript on PrimeKG-CL. We address each major comment below and will incorporate revisions to improve clarity and completeness.

read point-by-point responses

Referee: The benchmark relies on only two snapshots, corresponding to a single transition from June 2021 to July 2023. This setup does not capture cumulative effects, repeated updates, or varying change magnitudes typical in real biomedical KGs with independent ontology cycles. Consequently, the observed strong decoder-CL strategy interactions and the DistMult-specific separation between persistent and deprecated knowledge may be artifacts of this one-step regime rather than general properties.

Authors: We acknowledge that the use of only two snapshots captures a single transition rather than multi-step cumulative evolution. This design was necessitated by the availability of temporally aligned, high-quality releases from the nine source databases; additional intermediate snapshots with consistent schema and coverage are not publicly available for the full PrimeKG construction. The observed decoder-CL interactions hold consistently across three distinct tasks (relationship prediction, entity classification, and KGQA) and multiple metrics, which we interpret as evidence of robustness rather than an artifact of the one-step regime. Nevertheless, we agree this is a genuine limitation of the current benchmark. In the revised manuscript we will add an explicit limitations subsection in the discussion that quantifies the scale of the single transition (5.83 M added / 889 K removed edges) and outlines concrete directions for extending the benchmark with future data releases. revision: partial
Referee: Full details on hyperparameter choices, statistical significance testing across runs, and the precise construction of the 10 entity-type-grouped tasks with per-task persistent/added/removed stratification are not elaborated, which is load-bearing for assessing whether the reported decoder-strategy interactions are robust.

Authors: We apologize for the insufficient detail in the original submission. In the revised version we will expand Section 4 (Experiments) with: (i) complete hyperparameter grids and selection criteria for all four KGE decoders and six CL strategies, (ii) the exact statistical procedure (five random seeds, paired t-tests with reported p-values and standard deviations), and (iii) a step-by-step description of how the 10 entity-type-grouped tasks were derived, including the per-task rules for persistent/added/removed test stratification. These additions will be accompanied by new tables in the appendix. revision: yes

Circularity Check

0 steps flagged

Empirical benchmark release with no derivation chain or fitted predictions

full rationale

The paper introduces a new continual graph learning benchmark (PrimeKG-CL) constructed from two real temporal snapshots of biomedical KGs, defines 10 entity-type-grouped tasks with persistent/added/removed stratification, and reports empirical performance of six CL strategies across four KGE decoders plus additional baselines. No equations, first-principles derivations, or parameter-fitting steps are presented whose outputs are then relabeled as predictions. The central claims (decoder-CL interactions and DistMult-specific separation) are direct observations from the released experiments rather than reductions to self-defined quantities or self-citations. Self-citations, if any, are not load-bearing for any claimed result. The work is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The paper introduces an empirical benchmark rather than a theoretical model, so it carries no free parameters, ad-hoc axioms, or invented entities beyond the standard assumption that the chosen biomedical databases and snapshots are representative.

pith-pipeline@v0.9.0 · 5691 in / 1052 out tokens · 23211 ms · 2026-05-12T04:32:33.086004+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages · 2 internal anchors

[1]

Building a knowledge graph to enable precision medicine.Scientific Data, 10(1):67, 2023

Payal Chandak, Kexin Huang, and Marinka Zitnik. Building a knowledge graph to enable precision medicine.Scientific Data, 10(1):67, 2023

work page 2023
[2]

A foundation model for clinician-centered drug repurposing.Nature Medicine, 30(12):3601–3613, 2024

Kexin Huang, Payal Chandak, Qianwen Wang, Shreyas Haber, and Marinka Zitnik. A foundation model for clinician-centered drug repurposing.Nature Medicine, 30(12):3601–3613, 2024

work page 2024
[3]

Continual learning of knowledge graph embeddings.IEEE Robotics and Automation Letters, 6(2):1128–1135, 2021

Angel Daruna, Mehul Gupta, Mohan Sridharan, and Sonia Chernova. Continual learning of knowledge graph embeddings.IEEE Robotics and Automation Letters, 6(2):1128–1135, 2021

work page 2021
[4]

Overcoming catastrophic forgetting in neural networks.Proceedings of the National Academy of Sciences, 114(13):3521–3526, 2017

James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks.Proceedings of the National Academy of Sciences, 114(13):3521–3526, 2017

work page 2017
[5]

Experience replay for continual learning

David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy P Lillicrap, and Gregory Wayne. Experience replay for continual learning. InAdvances in Neural Information Processing Systems, volume 32, 2019

work page 2019
[6]

Lifelong embedding learning and transfer for growing knowledge graphs

Yuanning Cui, Yuxin Wang, Zequn Sun, Wenqiang Liu, Yiqiao Jiang, Kexin Han, and Wei Hu. Lifelong embedding learning and transfer for growing knowledge graphs. InProceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 4218–4226, 2023. 10

work page 2023
[7]

Towards continual knowledge graph embedding via incremental distillation

Jiajun Liu, Wenjun Ke, Peng Wang, Ziyu Shang, Jinhua Gao, Guozheng Li, Ke Ji, and Yanhe Liu. Towards continual knowledge graph embedding via incremental distillation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 8759–8768, 2024

work page 2024
[8]

Extended-connectivity fingerprints.Journal of Chemical Information and Modeling, 50(5):742–754, 2010

David Rogers and Mathew Hahn. Extended-connectivity fingerprints.Journal of Chemical Information and Modeling, 50(5):742–754, 2010

work page 2010
[9]

Continual learning through synaptic intelligence

Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In Proceedings of the 34th International Conference on Machine Learning, pages 3987–3995. PMLR, 2017

work page 2017
[10]

Distilling the Knowledge in a Neural Network

Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015

work page internal anchor Pith review Pith/arXiv arXiv 2015
[11]

Online continual learning with maximally interfered retrieval

Rahaf Aljundi, Lucas Caccia, Eugene Belilovsky, Massimo Caccia, Min Lin, Laurent Charlin, and Tinne Tuytelaars. Online continual learning with maximally interfered retrieval. InAdvances in Neural Information Processing Systems, volume 32, 2019

work page 2019
[12]

RotatE: Knowledge graph embedding by relational rotation in complex space

Zhiqing Sun, Zhi-Hong Deng, Jian-Yun Nie, and Jian Tang. RotatE: Knowledge graph embedding by relational rotation in complex space. InProceedings of the 7th International Conference on Learning Representations, 2019

work page 2019
[13]

Learning sequence encoders for temporal knowledge graph completion

Alberto García-Durán, Sebastijan Duman ˇci´c, and Mathias Niepert. Learning sequence encoders for temporal knowledge graph completion. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4816–4821, 2018

work page 2018
[14]

When Modalities Remember: Continual Learning for Multimodal Knowledge Graphs

Linyu Li, Zhi Jin, Yichi Zhang, Dongming Jin, Yuanpeng He, Haoran Duan, Gadeng Luosang, and Nyima Tashi. When modalities remember: Continual learning for multimodal knowledge graphs.arXiv preprint arXiv:2604.02778, 2026

work page internal anchor Pith review Pith/arXiv arXiv 2026
[15]

Translat- ing embeddings for modeling multi-relational data

Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. Translat- ing embeddings for modeling multi-relational data. InAdvances in Neural Information Processing Systems, volume 26, 2013

work page 2013
[16]

Embedding entities and relations for learning and inference in knowledge bases

Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. Embedding entities and relations for learning and inference in knowledge bases. InProceedings of the 3rd International Conference on Learning Representations, 2015

work page 2015
[17]

Fast and continual knowledge graph embedding via incremental LoRA

Jiajun Liu, Wenjun Ke, Peng Wang, Jiahao Wang, Jinhua Gao, Ziyu Shang, Guozheng Li, Zijie Xu, Ke Ji, and Yining Li. Fast and continual knowledge graph embedding via incremental LoRA. InProceedings of the 33rd International Joint Conference on Artificial Intelligence, pages 2159–2167, 2024

work page 2024
[18]

From knowledge forgetting to accumulation: Evolutionary relation path passing for lifelong knowledge graph embedding

Jing Yang, Xinfa Jiang, Xiaowen Jiang, Yuan Gao, Laurence T Yang, Shaojun Zou, and Shundong Yang. From knowledge forgetting to accumulation: Evolutionary relation path passing for lifelong knowledge graph embedding. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1197–1206, 2025

work page 2025
[19]

SAGE: Scale- aware gradual evolution for continual knowledge graph embedding

Yifei Li, Lingling Zhang, Hang Yan, Tianzhe Zhao, Zihan Ma, Muye Huang, and Jun Liu. SAGE: Scale- aware gradual evolution for continual knowledge graph embedding. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2025

work page 2025
[20]

DebiasedKGE: Towards mitigating spurious forgetting in continual knowledge graph embedding

Junlin Zhu, Bo Fu, and Guiduo Duan. DebiasedKGE: Towards mitigating spurious forgetting in continual knowledge graph embedding. InProceedings of the 34th ACM International Conference on Information and Knowledge Management (CIKM), 2025

work page 2025
[21]

Elastic weight consolidation for knowledge graph continual learning: An empirical evaluation.arXiv preprint arXiv:2512.01890, 2025

Gaganpreet Jhajj and Fuhua Lin. Elastic weight consolidation for knowledge graph continual learning: An empirical evaluation.arXiv preprint arXiv:2512.01890, 2025

work page arXiv 2025
[22]

Rethinking continual knowledge graph embedding: Benchmarks and analysis

Tianzhe Zhao, Jiaoyan Chen, Yanchi Ru, Qika Lin, Yuxia Geng, Haiping Zhu, Yudai Pan, and Jun Liu. Rethinking continual knowledge graph embedding: Benchmarks and analysis. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2025

work page 2025
[23]

Cglb: Benchmark tasks for continual graph learning

Xikun Zhang, Dongjin Song, and Dacheng Tao. Cglb: Benchmark tasks for continual graph learning. In Advances in Neural Information Processing Systems, volume 35, pages 13006–13021, 2022

work page 2022
[24]

Systematic integration of biomedical knowledge prioritizes drugs for repurposing.eLife, 6:e26726, 2017

Daniel Scott Himmelstein, Antoine Lizee, Christine Hessler, Leo Brueggeman, Sabrina L Chen, Dexter Hadley, Ari Green, Pouya Khankhanian, and Sergio E Baranzini. Systematic integration of biomedical knowledge prioritizes drugs for repurposing.eLife, 6:e26726, 2017. 11

work page 2017
[25]

Multimodal contrastive representation learning in augmented biomedical knowledge graphs.arXiv preprint arXiv:2501.01644, 2025

Tien Dang, Viet Thanh Duy Nguyen, Minh Tuan Le, and Truong-Son Hy. Multimodal contrastive representation learning in augmented biomedical knowledge graphs.arXiv preprint arXiv:2501.01644, 2025

work page arXiv 2025
[26]

Bridging the gap: Generating a comprehensive biomedical knowledge graph question answering dataset

Xi Yan, Patrick Westphal, Jan Seliger, and Ricardo Usbeck. Bridging the gap: Generating a comprehensive biomedical knowledge graph question answering dataset. InProceedings of the 27th European Conference on Artificial Intelligence (ECAI), volume 392 ofFrontiers in Artificial Intelligence and Applications, pages 1198–1205, 2024

work page 2024
[27]

MMKGR: Multi-hop multi-modal knowledge graph reasoning

Shangfei Zheng, Weiqing Wang, Jianfeng Qu, Hongzhi Yin, Wei Chen, and Lei Zhao. MMKGR: Multi-hop multi-modal knowledge graph reasoning. InProceedings of the 39th IEEE International Conference on Data Engineering (ICDE), pages 96–109, 2023

work page 2023
[28]

Multi-modal contrastive representation learning for entity alignment

Zhenxi Lin, Ziheng Zhang, Meng Wang, Yinghui Shi, Xian Wu, and Yefeng Zheng. Multi-modal contrastive representation learning for entity alignment. InProceedings of the 29th International Conference on Computational Linguistics, pages 2572–2584, 2022

work page 2022
[29]

Mose: Modality split and ensemble for multimodal knowledge graph completion

Yu Zhao, Xiangrui Cai, Yike Wu, Haiwei Zhang, Ying Zhang, Guoqing Zhao, and Ning Jiang. Mose: Modality split and ensemble for multimodal knowledge graph completion. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10527–10536, 2022

work page 2022
[30]

Balanced multimodal learning via on-the-fly gradient modulation

Xiaokang Peng, Yake Wei, Andong Deng, Dong Wang, and Di Hu. Balanced multimodal learning via on-the-fly gradient modulation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8238–8247, 2022

work page 2022
[31]

Characterizing and overcoming the greedy nature of learning in multi-modal deep neural networks

Nan Wu, Stanisław Jastrzebski, Kyunghyun Cho, and Krzysztof J Geras. Characterizing and overcoming the greedy nature of learning in multi-modal deep neural networks. InProceedings of the 39th International Conference on Machine Learning, pages 24043–24055. PMLR, 2022

work page 2022
[32]

Unifying large language models and knowledge graphs: A roadmap.IEEE Transactions on Knowledge and Data Engineering, 36(7):3580–3599, 2024

Shirui Pan, Linhao Luo, Yufei Wang, Chen Chen, Jiapu Wang, and Xindong Wu. Unifying large language models and knowledge graphs: A roadmap.IEEE Transactions on Knowledge and Data Engineering, 36(7):3580–3599, 2024

work page 2024
[33]

Retrieval-augmented generation for knowledge- intensive NLP tasks

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge- intensive NLP tasks. InAdvances in Neural Information Processing Systems, volume 33, pages 9459–9474, 2020

work page 2020
[34]

Three scenarios for continual learning

Gido M van de Ven and Andreas S Tolias. Three scenarios for continual learning.arXiv preprint arXiv:1904.07734, 2019

work page Pith review arXiv 1904
[35]

Domain-specific language model pretraining for biomedical natural language processing.ACM Transactions on Computing for Healthcare, 3(1):1–23, 2021

Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong Liu, Tristan Naumann, Jianfeng Gao, and Hoifung Poon. Domain-specific language model pretraining for biomedical natural language processing.ACM Transactions on Computing for Healthcare, 3(1):1–23, 2021

work page 2021
[36]

Modeling relational data with graph convolutional networks

Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. Modeling relational data with graph convolutional networks. InThe Semantic Web: 15th International Conference (ESWC 2018), pages 593–607. Springer, 2018

work page 2018
[37]

Complex embeddings for simple link prediction

Théo Trouillon, Johannes Welbl, Sebastian Riedel, Éric Gaussier, and Guillaume Bouchard. Complex embeddings for simple link prediction. InProceedings of the 33rd International Conference on Machine Learning, pages 2071–2080, 2016

work page 2071
[38]

KG-FIT: Knowledge graph fine-tuning upon open-world knowledge

Pengcheng Jiang, Lang Cao, Cao Xiao, Parminder Bhatia, Jimeng Sun, and Jiawei Han. KG-FIT: Knowledge graph fine-tuning upon open-world knowledge. InAdvances in Neural Information Processing Systems, volume 37, 2024

work page 2024
[39]

You CAN teach an old dog new tricks! on training knowledge graph embeddings

Daniel Ruffinelli, Samuel Broscheit, and Rainer Gemulla. You CAN teach an old dog new tricks! on training knowledge graph embeddings. InProceedings of the 8th International Conference on Learning Representations, 2020

work page 2020
[40]

Gradient surgery for multi-task learning

Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. InAdvances in Neural Information Processing Systems, volume 33, pages 5824–5836, 2020

work page 2020
[41]

forgetting

Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. InInternational Conference on Learning Representations, 2019. 12 PrimeKG-CL: A Continual Graph Learning Benchmark on Evolving Biomedical Knowledge Graphs Supplementary M...

work page 2019