pith. sign in

arxiv: 2507.03122 · v2 · pith:DSN4FMZMnew · submitted 2025-07-03 · 💻 cs.IR · cs.CL· cs.LG

Federated Learning for ICD Classification with Lightweight Models and Pretrained Embeddings

Binbin Xu , G\'erard Dray This is my paper

Pith reviewed 2026-05-22 00:51 UTC · model grok-4.3

classification 💻 cs.IR cs.CLcs.LG
keywords federated learningICD classificationclinical notespretrained embeddingsmultilayer perceptronprivacy-preservingMIMIC-IV
0
0 comments X

The pith

Embedding quality outweighs classifier complexity for ICD code prediction, and federated learning matches centralized results under balanced conditions.

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

The paper examines federated learning for assigning multiple ICD codes to clinical notes drawn from the MIMIC-IV dataset. It replaces large fine-tuned language models with frozen public embeddings fed into small multilayer perceptron heads. Experiments across six embedding models and three MLP sizes for both ICD-9 and ICD-10 show that the choice of embedding has a stronger effect on micro and macro F1 than the depth or width of the classifier. Federated averaging produces scores close to centralized training when data is evenly partitioned and label distributions match. The resulting models remain orders of magnitude smaller than current state-of-the-art systems while still delivering usable performance.

Core claim

Embedding quality substantially outweighs classifier complexity in determining predictive performance for multi-label ICD code classification, and federated learning can closely match centralized results in idealized conditions of even data splits and no communication failures.

What carries the argument

Frozen pretrained text embeddings paired with lightweight multilayer perceptron classifiers trained via federated averaging on distributed clinical notes.

If this is right

  • Small MLP heads on strong frozen embeddings reach competitive F1 scores for both ICD-9 and ICD-10 without end-to-end fine-tuning.
  • Federated averaging reproduces centralized performance when every site receives a balanced random subset of the same data distribution.
  • Performance remains stable when the same pipeline is rerun on ten different stratified splits.
  • Switching the embedding model changes results more than switching among the three tested MLP architectures.

Where Pith is reading between the lines

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

  • Hospitals could pool predictive power without moving raw notes, provided future work solves uneven data volumes and differing local coding practices.
  • The same lightweight pattern may transfer to other privacy-sensitive clinical tasks such as outcome prediction or billing validation.
  • Adding modest local adaptation layers at each site could close remaining gaps once real-world heterogeneity is introduced.

Load-bearing premise

Data is evenly split across sites with identical label distributions and no communication failures or site-specific shifts.

What would settle it

Repeat the federated experiments on partitions with strong label imbalance or site-specific coding preferences and measure whether the performance gap to the centralized baseline widens substantially.

Figures

Figures reproduced from arXiv: 2507.03122 by Binbin Xu, G\'erard Dray.

Figure 1
Figure 1. Figure 1: Overview of the proposed pipeline. Raw clinical text is encoded into dense embeddings using open-source models, followed by centralized or federated training of lightweight classifiers for multi-label ICD coding. 1https://huggingface.co/spaces/mteb/leaderboard [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: shows the relationship between input token length and inference time (in milliseconds) for the six selected embedding models, evaluated separately on discharge notes labeled with ICD-9 and ICD￾10 codes. As expected, inference time generally increases with token count, following an approximately linear or sublinear trend for most models [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Models architectural overview. The wide left blocks represent the precomputed embedding vectors. 2.4 Federated Learning Framework To implement federated training of this multi-label classification task, we evaluated several open-source FL libraries, with a focus on simplicity, cross-platform support, and performance: • Flower2 [30], is a lightweight, framework-independent FL platform designed for both rese… view at source ↗
Figure 4
Figure 4. Figure 4: Schematic representation of the federated learning framework used in this study, involving 20 simulated clients (nodes), each operating on isolated data partitions and participating in model updates through a central coordinating server. 2.5 Experimental Setup Experiments: Centralized & Federated training Two experimental training paradigms in this study are considered. In the centralized training setting,… view at source ↗
read the original abstract

This study investigates the feasibility and performance of federated learning (FL) for multi-label ICD code classification using clinical notes from the MIMIC-IV dataset. Unlike previous approaches that rely on centralized training or fine-tuned large language models, we propose a lightweight and scalable pipeline combining frozen text embeddings with simple multilayer perceptron (MLP) classifiers. This design offers a privacy-preserving and deployment-efficient alternative for clinical NLP applications, particularly suited to distributed healthcare settings. Extensive experiments across both centralized and federated configurations were conducted, testing six publicly available embedding models from Massive Text Embedding Benchmark leaderboard and three MLP classifier architectures under two medical coding (ICD-9 and ICD-10). Additionally, ablation studies over ten random stratified splits assess performance stability. Results show that embedding quality substantially outweighs classifier complexity in determining predictive performance, and that federated learning can closely match centralized results in idealized conditions. While the models are orders of magnitude smaller than state-of-the-art architectures and achieved competitive micro and macro F1 scores, limitations remain including the lack of end-to-end training and the simplified FL assumptions. Nevertheless, this work demonstrates a viable way toward scalable, privacy-conscious medical coding systems and offers a step toward for future research into federated, domain-adaptive clinical AI.

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.

Referee Report

3 major / 3 minor

Summary. The paper proposes a lightweight federated learning pipeline for multi-label ICD code classification on MIMIC-IV clinical notes. It freezes six pretrained text embeddings from the MTEB leaderboard, pairs them with three simple MLP classifier architectures, and evaluates under both centralized and federated training for ICD-9 and ICD-10 coding across ten stratified random splits. The central claims are that embedding quality drives larger performance differences than classifier depth or width, and that idealized federated learning (even splits, no failures or label shift) can closely match centralized results while remaining orders of magnitude smaller than end-to-end LLMs.

Significance. If the empirical comparisons hold, the work supplies a practical, privacy-preserving baseline for distributed clinical coding that avoids fine-tuning large models. The ablation design directly quantifies the relative impact of embedding choice versus MLP complexity and demonstrates competitive micro/macro F1 under the stated idealized FL assumptions, which is useful for resource-constrained healthcare deployments.

major comments (3)
  1. [§4.2, Table 3] §4.2 and Table 3: the claim that 'embedding quality substantially outweighs classifier complexity' rests on point estimates without reported standard deviations or statistical tests across the ten splits; the largest observed gap between embeddings could be within noise for some MLP configurations.
  2. [§3.3] §3.3: the federated setup assumes perfectly balanced client data partitions and identical label distributions; no ablation or sensitivity experiment is shown for heterogeneous splits or label shift, which directly limits the scope of the 'closely match centralized results' conclusion.
  3. [§4.1] §4.1: the methods section does not specify the exact client participation rate, number of communication rounds, or aggregation details (e.g., FedAvg vs. other variants) used in the reported FL runs, making it impossible to reproduce the exact centralized-to-FL gap.
minor comments (3)
  1. [Figure 1, §2.2] Figure 1 caption and §2.2: clarify whether the embedding models are used in their original form or with any domain-specific preprocessing of clinical notes.
  2. [Table 1] Table 1: add the exact parameter counts for each of the three MLP architectures to support the 'lightweight' claim.
  3. [§5] §5: the limitations paragraph mentions lack of end-to-end training but does not discuss potential negative transfer when embeddings are frozen on out-of-domain clinical text.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment below and have revised the manuscript accordingly to improve clarity, reproducibility, and the strength of our empirical claims.

read point-by-point responses

  1. Referee: [§4.2, Table 3] §4.2 and Table 3: the claim that 'embedding quality substantially outweighs classifier complexity' rests on point estimates without reported standard deviations or statistical tests across the ten splits; the largest observed gap between embeddings could be within noise for some MLP configurations.

    Authors: We agree that reporting variability strengthens the claim. In the revised manuscript we will update Table 3 to show mean and standard deviation of micro- and macro-F1 across the ten stratified splits for every embedding–MLP pair. While we did not conduct formal pairwise statistical tests, the consistent ranking of embeddings across all splits and classifier depths supports the conclusion that embedding quality produces larger and more stable differences than changes in MLP width or depth. The added statistics will allow readers to judge whether observed gaps exceed observed variability. revision: yes

  2. Referee: [§3.3] §3.3: the federated setup assumes perfectly balanced client data partitions and identical label distributions; no ablation or sensitivity experiment is shown for heterogeneous splits or label shift, which directly limits the scope of the 'closely match centralized results' conclusion.

    Authors: Our experiments deliberately use idealized, balanced partitions to isolate the effect of the federated training procedure itself under controlled conditions. The manuscript already lists simplified FL assumptions among its limitations. We will expand the text in §3.3 and the limitations paragraph to state explicitly that the reported centralized-to-FL gaps hold only under even splits and identical label distributions, and that real-world label shift or non-IID partitions may widen the gap. We view this as a scope clarification rather than a new experimental ablation at the minor-revision stage. revision: partial

  3. Referee: [§4.1] §4.1: the methods section does not specify the exact client participation rate, number of communication rounds, or aggregation details (e.g., FedAvg vs. other variants) used in the reported FL runs, making it impossible to reproduce the exact centralized-to-FL gap.

    Authors: We thank the referee for noting this omission. The revised §4.1 will state that every client participates in each round (100 % participation), that training proceeds for 50 communication rounds, and that the FedAvg algorithm is used for server-side aggregation. These exact settings were employed for all federated runs reported in the paper and will now be documented for full reproducibility. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper reports an empirical comparison of six pretrained embeddings against three MLP architectures for multi-label ICD classification on MIMIC-IV, measuring micro/macro F1 on held-out splits across ten stratified random partitions in both centralized and federated settings. All performance numbers are obtained by direct evaluation on test data rather than by any derivation, fitting step, or self-referential prediction; the idealized FL assumptions (even splits, no failures or label shift) are explicitly scoped and listed as limitations. No equations, uniqueness theorems, or ansatzes appear, and no self-citations are invoked to justify core claims. The central finding that embedding quality dominates classifier complexity is therefore a measured experimental outcome, not a reduction to the paper's own inputs.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on the assumption that MIMIC-IV notes are representative of real clinical text and that the Massive Text Embedding Benchmark models transfer without domain adaptation. No new entities are postulated.

free parameters (2)
  • MLP hidden size and depth
    Three architectures tested; sizes chosen to keep models lightweight.
  • Federated round count and client participation rate
    Not specified in abstract but required for any FL run.
axioms (2)
  • domain assumption Data can be partitioned across simulated clients without label or feature shift that would break convergence.
    Invoked when claiming FL matches centralized performance under idealized conditions.
  • domain assumption Frozen embeddings require no further gradient updates for the target task.
    Stated as part of the lightweight pipeline design.

pith-pipeline@v0.9.0 · 5751 in / 1466 out tokens · 38142 ms · 2026-05-22T00:51:59.728954+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages · 6 internal anchors

  1. [1]

    International classification of diseases - icd,

    World Health Organization et al., “International classification of diseases - icd,” 2009. 18

    work page 2009

  2. [2]

    Diagnosis code assignment: models and evaluation metrics,

    A. Perotte, R. Pivovarov, K. Natarajan, N. Weiskopf, F. Wood, and N. Elhadad, “Diagnosis code assignment: models and evaluation metrics,” Journal of the American Medical Informatics Association, vol. 21, pp. 231–237, 12 2013

    work page 2013

  3. [3]

    Ehr coding with multi-scale feature attention and struc- tured knowledge graph propagation,

    X. Xie, Y. Xiong, P. S. Yu, and Y. Zhu, “Ehr coding with multi-scale feature attention and struc- tured knowledge graph propagation,” in Proceedings of the 28th ACM International Conference on Information and Knowledge Management , CIKM ’19, (New York, NY, USA), p. 649–658, Associ- ation for Computing Machinery, 2019

    work page 2019

  4. [4]

    Explainable Prediction of Medical Codes from Clinical Text

    J. Mullenbach, S. Wiegreffe, J. Duke, J. Sun, and J. Eisenstein, “Explainable prediction of medical codes from clinical text,” arXiv preprint arXiv:1802.05695 , 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  5. [5]

    Multi-label classification of patient notes: Case study on icd code assignment.,

    T. Baumel, J. Nassour-Kassis, R. Cohen, M. Elhadad, and N. Elhadad, “Multi-label classification of patient notes: Case study on icd code assignment.,” in AAAI Workshops, pp. 409–416, 2018

    work page 2018

  6. [6]

    Automated icd-9 coding via a deep learning approach,

    M. Li, Z. Fei, M. Zeng, F. Wu, Y. Li, Y. Pan, and J. Wang, “Automated icd-9 coding via a deep learning approach,” IEEE/ACM Transactions on Computational Biology and Bioinformatics, vol. 16, pp. 1193–1202, July 2019

    work page 2019

  7. [7]

    Joint embedding of words and labels for text classification,

    G. Wang, C. Li, W. Wang, Y. Zhang, D. Shen, X. Zhang, R. Henao, and L. Carin, “Joint embedding of words and labels for text classification,” in Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (I. Gurevych and Y. Miyao, eds.), (Melbourne, Australia), pp. 2321–2331, Association for Computationa...

    work page 2018

  8. [8]

    Automatic icd- 10 multi-class classification of cause of death from plaintext autopsy reports through expert-driven feature selection,

    G. Mujtaba, L. Shuib, R. G. Raj, R. Rajandram, K. Shaikh, and M. A. Al-Garadi, “Automatic icd- 10 multi-class classification of cause of death from plaintext autopsy reports through expert-driven feature selection,” PLOS ONE, vol. 12, pp. 1–27, 02 2017

    work page 2017

  9. [9]

    Automatic classification of diseases from free-text death certificates for real-time surveillance,

    B. Koopman, S. Karimi, A. Nguyen, R. McGuire, D. Muscatello, M. Kemp, D. Truran, M. Zhang, and S. Thackway, “Automatic classification of diseases from free-text death certificates for real-time surveillance,” BMC Medical Informatics and Decision Making , vol. 15, no. 1, pp. 53–, 2015

    work page 2015

  10. [10]

    Automatic matching of icd-10 codes to diagnoses in discharge letters,

    S. Boytcheva, “Automatic matching of icd-10 codes to diagnoses in discharge letters,” inProceedings of the second workshop on biomedical natural language processing , pp. 11–18, 2011

    work page 2011

  11. [11]

    Chen, T.-L

    P.-F. Chen, T.-L. He, S.-C. Lin, Y.-C. Chu, C.-T. Kuo, F. Lai, S.-M. Wang, W.-X. Zhu, K.-C. Chen, L.-C. Kuo, F.-M. Hung, Y.-C. Lin, I.-C. Tsai, C.-H. Chiu, S.-C. Chang, and C.-Y. Yang, “Training a deep contextualized language model for international classification of diseases, 10th revision classification via federated learning: Model development and vali...

    work page 2022

  12. [12]

    PLM-ICD: Automatic ICD coding with pretrained lan- guage models,

    C.-W. Huang, S.-C. Tsai, and Y.-N. Chen, “PLM-ICD: Automatic ICD coding with pretrained lan- guage models,” in Proceedings of the 4th Clinical Natural Language Processing Workshop (T. Nau- mann, S. Bethard, K. Roberts, and A. Rumshisky, eds.), (Seattle, WA), pp. 10–20, Association for Computational Linguistics, July 2022

    work page 2022

  13. [13]

    Pre-Training a neural language model improves the sample efficiency of an emergency room classification model,

    B. Xu, C. Gil-Jardin´ e, F. Thiessard, E. Tellier, M. Avalos, and E. Lagarde, “Pre-Training a neural language model improves the sample efficiency of an emergency room classification model,” in The 33rd Florida Artificial Intelligence Research Society Conference , 2020

    work page 2020

  14. [14]

    Benchmarking pysyft federated learning framework on mimic-iii dataset,

    A. Budrionis, M. Miara, P. Miara, S. Wilk, and J. G. Bellika, “Benchmarking pysyft federated learning framework on mimic-iii dataset,” IEEE Access, vol. 9, pp. 116869–116878, 2021. 19

    work page 2021

  15. [15]

    Flicu: A federated learning workflow for intensive care unit mortality prediction,

    L. Mondrejevski, I. Miliou, A. Montanino, D. Pitts, J. Hollm´ en, and P. Papapetrou, “Flicu: A federated learning workflow for intensive care unit mortality prediction,” in 2022 IEEE 35th In- ternational Symposium on Computer-Based Medical Systems (CBMS) , pp. 32–37, July 2022

    work page 2022

  16. [16]

    Exploratory analysis of federated learning methods with differential privacy on mimic-iii,

    A. N. Horvath, M. Berchier, F. Nooralahzadeh, A. Allam, and M. Krauthammer, “Exploratory analysis of federated learning methods with differential privacy on mimic-iii,” arXiv preprint arXiv:2302.04208, 2023

    work page arXiv 2023

  17. [17]

    Mimic-iv-note: Deidentified free-text clinical notes,

    A. Johnson, T. Pollard, S. Horng, L. A. Celi, and R. Mark, “Mimic-iv-note: Deidentified free-text clinical notes,” 2023

    work page 2023

  18. [18]

    Physiobank, physiotoolkit, and physionet: components of a new research resource for complex physiologic signals,

    A. L. Goldberger, L. A. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E. Stanley, “Physiobank, physiotoolkit, and physionet: components of a new research resource for complex physiologic signals,”circulation, vol. 101, no. 23, pp. e215–e220, 2000

    work page 2000

  19. [19]

    MTEB: Massive Text Embedding Benchmark

    N. Muennighoff, N. Tazi, L. Magne, and N. Reimers, “Mteb: Massive text embedding benchmark,” arXiv preprint arXiv:2210.07316 , 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  20. [20]

    Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models

    Y. Zhang, M. Li, D. Long, X. Zhang, H. Lin, B. Yang, P. Xie, A. Yang, D. Liu, J. Lin, et al. , “Qwen3 embedding: Advancing text embedding and reranking through foundation models,” arXiv preprint arXiv:2506.05176, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  21. [21]

    Qwen3 Technical Report

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, et al. , “Qwen3 technical report,” arXiv preprint arXiv:2505.09388 , 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  22. [22]

    Towards General Text Embeddings with Multi-stage Contrastive Learning

    Z. Li, X. Zhang, Y. Zhang, D. Long, P. Xie, and M. Zhang, “Towards general text embeddings with multi-stage contrastive learning,” arXiv preprint arXiv:2308.03281 , 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  23. [23]

    Qwen2 technical report,

    A. Yang, B. Yang, B. Hui, B. Zheng, B. Yu, C. Zhou, C. Li, C. Li, D. Liu, F. Huang, G. Dong, H. Wei, H. Lin, J. Tang, J. Wang, J. Yang, J. Tu, J. Zhang, J. Ma, J. Yang, J. Xu, J. Zhou, J. Bai, J. He, J. Lin, K. Dang, K. Lu, K. Chen, K. Yang, M. Li, M. Xue, N. Ni, P. Zhang, P. Wang, R. Peng, R. Men, R. Gao, R. Lin, S. Wang, S. Bai, S. Tan, T. Zhu, T. Li, T...

    work page 2024

  24. [24]

    mgte: Generalized long-context text representation and reranking models for multilingual text retrieval,

    X. Zhang, Y. Zhang, D. Long, W. Xie, Z. Dai, J. Tang, H. Lin, B. Yang, P. Xie, F. Huang, et al., “mgte: Generalized long-context text representation and reranking models for multilingual text retrieval,” in Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track, pp. 1393–1412, 2024

    work page 2024

  25. [25]

    Arctic-embed 2.0: Multilingual retrieval without compromise, 2024

    P. Yu, L. Merrick, G. Nuti, and D. Campos, “Arctic-embed 2.0: Multilingual retrieval without compromise,” arXiv preprint arXiv:2412.04506 , 2024

    work page arXiv 2024

  26. [26]

    Nomic embed: Training a reproducible long context text embedder,

    Z. Nussbaum, J. X. Morris, A. Mulyar, and B. Duderstadt, “Nomic embed: Training a reproducible long context text embedder,” Transactions on Machine Learning Research, 2025. Reproducibility Certification

    work page 2025

  27. [27]

    Automated medical coding on mimic-iii and mimic-iv: A critical review and replicability study,

    J. Edin, A. Junge, J. D. Havtorn, L. Borgholt, M. Maistro, T. Ruotsalo, and L. Maaløe, “Automated medical coding on mimic-iii and mimic-iv: A critical review and replicability study,” in Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’23, (New York, NY, USA), p. 2572–2582, Association...

    work page 2023

  28. [28]

    Icd coding from clinical text using multi-filter residual convolutional neural network,

    F. Li and H. Yu, “Icd coding from clinical text using multi-filter residual convolutional neural network,” Proceedings of the AAAI Conference on Artificial Intelligence , vol. 34, pp. 8180–8187, Apr. 2020

    work page 2020

  29. [29]

    A label attention model for icd coding from clinical text,

    T. Vu, D. Q. Nguyen, and A. Nguyen, “A label attention model for icd coding from clinical text,” in Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI’20, 2021

    work page 2021

  30. [30]

    Flower: A Friendly Federated Learning Research Framework

    D. J. Beutel, T. Topal, A. Mathur, X. Qiu, J. Fernandez-Marques, Y. Gao, L. Sani, K. H. Li, T. Parcollet, P. P. B. de Gusm˜ ao,et al., “Flower: A friendly federated learning research framework,” arXiv preprint arXiv:2007.14390 , 2020

    work page internal anchor Pith review arXiv 2007

  31. [31]

    Ziller, A

    A. Ziller, A. Trask, A. Lopardo, B. Szymkow, B. Wagner, E. Bluemke, J.-M. Nounahon, J. Passerat- Palmbach, K. Prakash, N. Rose, T. Ryffel, Z. N. Reza, and G. Kaissis, PySyft: A Library for Easy Federated Learning, pp. 111–139. Cham: Springer International Publishing, 2021

    work page 2021

  32. [32]

    Fedlab: A flexible federated learning framework,

    D. Zeng, S. Liang, X. Hu, H. Wang, and Z. Xu, “Fedlab: A flexible federated learning framework,” Journal of Machine Learning Research , vol. 24, no. 100, pp. 1–7, 2023

    work page 2023

  33. [33]

    Secureboost: A lossless federated learning framework,

    K. Cheng, T. Fan, Y. Jin, Y. Liu, T. Chen, D. Papadopoulos, and Q. Yang, “Secureboost: A lossless federated learning framework,” IEEE Intelligent Systems , vol. 36, pp. 87–98, Nov 2021

    work page 2021

  34. [34]

    Mimic- iv-icd: A new benchmark for extreme multilabel classification,

    T.-T. Nguyen, V. Schlegel, A. Kashyap, S. Winkler, S.-S. Huang, J.-J. Liu, and C.-J. Lin, “Mimic- iv-icd: A new benchmark for extreme multilabel classification,” arXiv preprint arXiv:2304.13998 , 2023

    work page arXiv 2023