Pith. sign in

REVIEW 3 major objections 5 minor 60 references

Proxy Model-Guided Reinforcement Learning for Client Selection in Federated Recommendation

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Proxy-predicted client losses can replace costly pre-round training in federated recommendation.

desk verdict A plausible and genuinely new combination—proxy-predicted per-client loss plus a staleness-aware RL reward—but the significance claim is unsupported and the proxy cold-start for never-selected clients is a real gap. read the letter →

arxiv 2508.10401 v1 pith:3TPAYP4I submitted 2025-08-14 cs.IR

Liang Qu , Jianxin Li , Wei Yuan , Penghui Ruan , Yuhui Shi , Hongzhi Yin This is my paper · ORCID
classification cs.IR
keywords federatedrecommendationclientselectionreinforcementlearningproxymodelembeddingstalenesslong-tailitemsNeuralCollaborativeFiltering
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

The paper is trying to establish that client selection in federated recommendation can be made both smarter and cheaper: a lightweight "proxy" branch inside each client's recommendation model can predict that client's training loss without running local training, and a reinforcement learning agent that consumes these predicted losses—plus a staleness penalty favoring long-tail item embeddings—picks better clients than random selection or generic federated-learning selectors. The claim is tested on three public datasets, with reported gains largest on the sparsest ones (e.g., HR@20 of 0.0576 vs 0.0494 for the best baseline on Fashion). If correct, contribution evaluation no longer scales with per-client training cost, and selection can explicitly counter the item-embedding staleness that hurts federated recommendation.

What carries the argument

ProxyNCF is the dual-branch model: a second, lightweight branch shares the user/item embeddings and MLP structure of NCF but is trained by MSE to predict the BPR loss ($\hat{\ell}_{u,v,j}$ vs $\ell_{u,v,j}$). It supplies the predicted losses used as the RL state. The second machinery is the staleness-aware reward, which subtracts a staleness penalty $\frac{1}{|V|}\sum_v \tau^{(t)}(v)/T$ from the validation-accuracy reward, giving the actor-critic agent an incentive to select clients that refresh long-tail embeddings.

What would settle it

Compare the proxy-predicted loss $\hat{L}_u$ against the true BPR loss computed on the same client after local training, for a held-out set of clients across training rounds; if the correlation for never-selected clients is near zero, the RL state is largely noise and the accuracy gap over random selection should shrink as the client pool grows.

Watch

Extended reading notes

Core claim

The paper claims that the two obstacles to non-random client selection in federated recommendation—the cost of evaluating thousands of clients and the sparsity of item-embedding updates—can be addressed by pairing a cheap prediction model with a purpose-built reinforcement learning reward. Each client runs ProxyNCF, an NCF model with an extra proxy branch. In every round the proxy branch outputs a predicted BPR loss $\hat{L}_u$ for that client; the server packs all predicted losses into the RL state, and an actor-critic agent samples the set of clients to train. The reward is $\lambda \cdot \mathrm{Acc}^{(t)} - (1-\lambda) \cdot \mathrm{Staleness}^{(t)}$, where staleness averages the rounds

Load-bearing premise

The load-bearing premise is that every client's proxy branch, including branches on clients that are never selected and therefore never trained, produces a trustworthy estimate of that client's training loss.

Editorial extensions

If this is right

  • Contribution estimation becomes a single inference pass per client per round instead of per-round local training, so selection cost grows slowly with client count.
  • Staleness-aware rewards give long-tail and rarely updated item embeddings a path to be refreshed, which is where random selection performs worst.
  • Because the proxy branch is decoupled from the recommendation objective, the selection policy can be attached to different embedding-based backbones without changing the federated training loop.
  • On the reported experiments, the accuracy gains over baselines are larger on sparse datasets than on the denser one, suggesting the method is most valuable where interactions are scarce.

Reading between the lines

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

  • Editorial inference: the RL state is built from proxy predictions for all clients, but only selected clients ever train their proxy branch; if predictions for never-selected clients are stale or uninformative, part of the state is noise and reported gains could erode as the client pool grows or the number of rounds shrinks.
  • Editorial inference: using predicted loss both as the selection signal and as a component of what the agent improves may create a feedback loop that favors clients whose proxies are optimistic rather than clients whose updates actually help the global model; a three-way comparison of selection by predicted loss, true loss, and random would isolate this.
  • The paper's own limitation section notes that it assumes static user data and that the proxy module adds communication overhead; extending the approach to streaming interactions is left as future work.
  • A natural calibration extension would be to have the server occasionally probe a small random holdout of clients, compute their true losses, and use those to rescale proxy predictions before the RL state is constructed.
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 / 5 minor

Summary. The paper addresses client selection in federated recommender systems (FedRSs). It proposes ProxyRL-FRS, which combines a client-side dual-branch model (ProxyNCF) with a server-side staleness-aware reinforcement learning agent. The NCF branch performs standard recommendation training, while the proxy branch predicts each client's BPR loss without requiring local training, thus avoiding the pre-round contribution-evaluation cost. The RL agent selects clients from the proxy-predicted losses, and the reward function trades off validation accuracy against item-embedding staleness. Experiments on MovieLens-1M, Fashion, and Video Games compare against random and non-random client-selection baselines. The paper claims that ProxyRL-FRS consistently achieves the highest HR@20 and NDCG@20 across all datasets with competitive training time, and ablations isolate the contributions of the proxy branch, staleness term, and accuracy term.

Significance. If the empirical claims are substantiated, this is a practical contribution to FedRS: proxy-estimated losses give a cheap communication- and computation-efficient selection signal, and the staleness-aware reward explicitly targets the long-tail embedding-update problem. The method is described in enough detail to reimplement, with clear equations and Algorithm 1. The w/o Proxy ablation in Table IV is informative: using true losses gives nearly the same accuracy but much higher training time, demonstrating the efficiency benefit of the proxy. The central accuracy claim, however, is currently supported only by single-run table entries; no statistical evidence is supplied, and the baseline comparisons mix different backbone models. These gaps mean the headline result is plausible but not yet established.

major comments (3)
  1. [Table III / Section V.E] The caption states 'All improvements are statistically significant with p < 0.05', but no variance, number of seeds, or significance test is reported anywhere. Every cell is a single number. Given that some differences are small (e.g., MovieLens-1M HR@20 0.4579 vs. 0.4473 for Pow-d), the claim cannot be verified. Please report mean±std over multiple runs and specify the test used.
  2. [Table III / Section V.B] The comparison conflates client-selection strategy with backbone recommendation model. Random-CS rows use different base models (FedMF, FedNCF, PFedRec, FedPerGNN, GPFedRec), and the non-random baselines also use their own backbones, while ProxyRL-FRS uses NCF. Thus Table III measures combined differences, not the effect of selection alone. A controlled comparison with the same backbone (e.g., FedNCF with random, Pow-d, FedFast, FedRL, and ProxyRL-FRS selectors) is needed to support the claim that the selection strategy, rather than the backbone, drives the gains.
  3. [Eq. (16) and Algorithm 1, lines 3-13] The RL state s(t) uses proxy-predicted losses for every client, but the proxy branch is trained only on selected clients with ground-truth losses. For clients never selected, the proxy branch receives no supervision and, because they do not receive the global item embeddings in Eq. (1), their local embeddings become stale. Their predicted losses are therefore outputs of untrained or outdated branches, i.e., essentially noise. In the Fashion dataset (29,858 clients), if |U+| is 100 and training runs 100 rounds, at most 10,000 client-round updates occur, so a large fraction of clients may never be selected. The w/o Proxy ablation (Table IV) does not address this, since it uses true losses for all clients. Please provide either a mechanism for updating or calibrating the proxy branches of unselected clients, or evidence (e.g., correlation between predicted and true losses on held-out client
minor comments (5)
  1. [Section V.D] The default values for |U+|, λ, staleness window T, and proxy-branch regression hyperparameters are not reported. These are needed to reproduce Table III and Figure 6.
  2. [Algorithm 1] Line 7 uses 'st' while Eq. (16) defines s(t). Please unify the notation.
  3. [Table III] The random-CS row lists 'PFedGNN', but Section V.B names the method 'FedPerGNN'. Please correct the inconsistency.
  4. [Section IV.A.2] Typo: 'never shared with the serve' should be 'never shared with the server'.
  5. [Figure 6] The figure reports cumulative training time, but the hardware, number of clients simulated per round, and parallelization details are absent. Please add these experimental conditions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the proxy branch is supervised by ground-truth losses and the RL reward uses external validation metrics; the central claim is empirical and not reduced to its inputs.

full rationale

ProxyRL-FRS's derivation chain is self-contained. The proxy branch (Eq. 14) is trained by MSE regression against ground-truth BPR losses (Eq. 15), so its predictions are anchored to real training targets rather than to the selection outcome or to the final accuracy metric. The RL state (Eq. 16) uses these proxy-estimated losses, but the reward (Eq. 17) is computed from validation HR/NDCG and server-side embedding staleness (Eq. 18), both external to the proxy's own predictions. The w/o Proxy ablation (Table IV) replaces the proxy with true losses and obtains similar accuracy, confirming that the proxy is a computational substitute rather than a renamed version of the target. The RL policy is optimized against this external reward, and the reported test metrics are held out, so no fitted parameter is being reported as a prediction. Self-citations in the related-work and reference sections are contextual and not load-bearing; no uniqueness theorem or ansatz is imported from co-authored prior work. The concern about untrained/stale proxy branches for never-selected clients is a robustness and scalability limitation, not a circularity: even if those predictions are noisy, the method's empirical gains could degrade, which is a correctness risk rather than a definitional reduction. The paper is appropriately benchmarked against external baselines, and its central claim is empirical. Therefore, no circular step is present.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The method's selection quality rests on the proxy branch's loss prediction (Eq. 14-15) being a valid contribution signal, on the RL agent being able to shape a useful policy from about 100 scalar-reward rounds, and on the server being able to reduce embedding staleness without per-client item visibility (Eq. 18). These are domain assumptions, not derived results. The free parameters are the reward trade-off lambda, the number of selected clients |U+|, the staleness window T, and the proxy training hyperparameters; none of their default values is stated for the main experiments.

free parameters (4)
  • lambda (accuracy-staleness reward trade-off) = not stated for main results; swept over {0, 0.2, 0.4, 0.6, 0.8, 1} on MovieLens-1M only
    Weights the two reward terms in Eq. 17; the value used in Table III is never reported.
  • number of selected clients |U+| = not stated for main results; tested {50, 100, 200, 500} on MovieLens-1M
    Determines per-round participation and thus staleness dynamics; the default for Table III is not given.
  • staleness normalization window T = not stated
    Normalizes the staleness term in Eq. 18; its value changes the reward scale and therefore the learned policy.
  • proxy branch regression hyperparameters = not stated
    Proxy branch is trained with MSE (Eq. 15); learning rate, batch size, local epochs, and negative sampling count are unreported.
assumptions (5)
  • domain assumption Predicted BPR loss from the proxy branch is an adequate measure of client contribution for selection.
    The state (Eq. 16) and the whole selection pipeline rest on this; it is validated only indirectly by the w/o Proxy ablation.
  • domain assumption The server can reduce item-embedding staleness through client selection without knowing per-client item membership.
    Privacy keeps membership local (Section IV-A2, Eq. 10), so the agent receives only aggregate staleness (Eq. 18) with no direct credit assignment from action to staleness reduction.
  • domain assumption Policy-gradient and TD updates over about 100 rounds of scalar reward can shape a useful selection policy.
    Algorithm 1 line 16 updates the RL agent once per round with a single reward signal; no convergence evidence for the agent itself is provided.
  • domain assumption Local user data is static during training.
    Stated as a limitation in Section VI.B; streaming data would change both the loss signal and the staleness computation.
  • standard math FedAvg aggregation (Eq. 19) and BPR loss (Eq. 13) are used unmodified and behave as in the literature.
    The global update and the local objective are standard components taken from prior work without modification.
invented entities (1)
  • Proxy branch of ProxyNCF (dual-branch module)
    purpose: Predicts each client's BPR training loss without full local training (Eq. 14) and forms the RL state (Eq. 16).
    A learned estimator validated only internally: the w/o Proxy ablation shows equal accuracy at 4.6x higher training time on Fashion, but no external benchmark of prediction quality is given, and the branch is never trained for never-selected clients.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Proxy Model-Guided Reinforcement Learning for Client Selection in Federated Recommendation." pith.science (2026). https://pith.science/paper/3TPAYP4I

@misc{pith2026250810401,
  author       = {Pith},
  title        = {Pith review of: Proxy Model-Guided Reinforcement Learning for Client Selection in Federated Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3TPAYP4I}},
  note         = {Machine review of arXiv:2508.10401}
}
read the original abstract

Federated recommender systems have emerged as a promising privacy-preserving paradigm, enabling personalized recommendation services without exposing users' raw data. By keeping data local and relying on a central server to coordinate training across distributed clients, FedRSs protect user privacy while collaboratively learning global models. However, most existing FedRS frameworks adopt fully random client selection strategy in each training round, overlooking the statistical heterogeneity of user data arising from diverse preferences and behavior patterns, thereby resulting in suboptimal model performance. While some client selection strategies have been proposed in the broader federated learning literature, these methods are typically designed for generic tasks and fail to address the unique challenges of recommendation scenarios, such as expensive contribution evaluation due to the large number of clients, and sparse updates resulting from long-tail item distributions. To bridge this gap, we propose ProxyRL-FRS, a proxy model-guided reinforcement learning framework tailored for client selection in federated recommendation. Specifically, we first introduce ProxyNCF, a dual-branch model deployed on each client, which augments standard Neural Collaborative Filtering with an additional proxy model branch that provides lightweight contribution estimation, thus eliminating the need for expensive per-round local training traditionally required to evaluate a client's contribution. Furthermore, we design a staleness-aware SA reinforcement learning agent that selects clients based on the proxy-estimated contribution, and is guided by a reward function balancing recommendation accuracy and embedding staleness, thereby enriching the update coverage of item embeddings. Experiments conducted on public recommendation datasets demonstrate the effectiveness of ProxyRL-FRS.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 48 canonical work pages

  1. [1]

    Deep neural networks for youtube recommendations,

    P. Covington, J. Adams, and E. Sargin, “Deep neural networks for youtube recommendations,” in Proceedings of the 10th ACM conference on recommender systems , 2016, pp. 191–198

    work page 2016

  2. [2]

    Amazon. com recommendations: Item-to-item collaborative filtering,

    G. Linden, B. Smith, and J. York, “Amazon. com recommendations: Item-to-item collaborative filtering,” IEEE Internet computing , vol. 7, no. 1, pp. 76–80, 2003

    work page 2003

  3. [3]

    A survey on federated learning,

    C. Zhang, Y . Xie, H. Bai, B. Yu, W. Li, and Y . Gao, “A survey on federated learning,” Knowledge-Based Systems , vol. 216, p. 106775, 2021

    2021

  4. [4]

    A Survey on Federated Recommendation Systems

    Z. Sun, Y . Xu, Y . Liu, W. He, Y . Jiang, F. Wu, and L. Cui, “A survey on federated recommendation systems,” arXiv preprint arXiv:2301.00767 , 2022

    work page Pith review arXiv 2022

  5. [5]

    Federated recommendation systems,

    L. Yang, B. Tan, V . W. Zheng, K. Chen, and Q. Yang, “Federated recommendation systems,” Federated Learning: Privacy and Incentive , pp. 225–239, 2020

    work page 2020

  6. [6]

    Federated recommenders: methods, challenges and future,

    Z. Alamgir, F. K. Khan, and S. Karim, “Federated recommenders: methods, challenges and future,” Cluster Computing, vol. 25, no. 6, pp. 4075–4096, 2022

    work page 2022

  7. [7]

    A comprehensive survey on privacy-preserving techniques in federated recommendation systems,

    M. Asad, S. Shaukat, E. Javanmardi, J. Nakazato, and M. Tsukada, “A comprehensive survey on privacy-preserving techniques in federated recommendation systems,” Applied Sciences , vol. 13, no. 10, p. 6201, 2023

    work page 2023

  8. [8]

    Recent advances and future challenges in federated recommender systems,

    M. Harasic, F.-S. Keese, D. Mattern, and A. Paschke, “Recent advances and future challenges in federated recommender systems,” International Journal of Data Science and Analytics , vol. 17, no. 4, pp. 337–357, 2024

    work page 2024

Show all 60 references
  1. [9]

    A survey on cross-user federated recommendation,

    E. Yang, Y . Xiong, W. Yuan, W. Pan, Q. Yang, and Z. Ming, “A survey on cross-user federated recommendation,” Science China Information Sciences, vol. 68, no. 4, pp. 1–26, 2025

    2025

  2. [10]

    Privacy-preserved and responsible recommenders: From conventional defense to federated learning and blockchain,

    W. Ali, X. Zhou, and J. Shao, “Privacy-preserved and responsible recommenders: From conventional defense to federated learning and blockchain,” ACM Comput. Surv. , vol. 57, no. 5, Jan. 2025. [Online]. Available: https://doi.org/10.1145/3708982

    2025 doi

  3. [11]

    On-device recommender systems: A comprehensive survey,

    H. Yin, L. Qu, T. Chen, W. Yuan, R. Zheng, J. Long, X. Xia, Y . Shi, and C. Zhang, “On-device recommender systems: A comprehensive survey,” arXiv preprint arXiv:2401.11441 , 2024

    2024 arXiv

  4. [12]

    Client selection in federated learning: Principles, challenges, and opportunities,

    L. Fu, H. Zhang, G. Gao, M. Zhang, and X. Liu, “Client selection in federated learning: Principles, challenges, and opportunities,” IEEE Internet of Things Journal , 2023

    2023

  5. [13]

    A review of client selection methods in federated learning,

    S. Mayhoub and T. M. Shami, “A review of client selection methods in federated learning,” Archives of Computational Methods in Engineering, vol. 31, no. 2, pp. 1129–1152, 2024

    2024

  6. [14]

    A comprehensive survey on client selection strategies in federated learning,

    J. Li, T. Chen, and S. Teng, “A comprehensive survey on client selection strategies in federated learning,” Computer Networks, p. 110663, 2024

    2024

  7. [15]

    Auction: Automated and quality-aware client selection framework for efficient federated learning,

    Y . Deng, F. Lyu, J. Ren, H. Wu, Y . Zhou, Y . Zhang, and X. Shen, “Auction: Automated and quality-aware client selection framework for efficient federated learning,” IEEE Transactions on Parallel and Dis- tributed Systems, vol. 33, no. 8, pp. 1996–2009, 2021

    1996

  8. [16]

    Client selection in federated learning based on gradients importance,

    O. Marnissi, H. E. Hammouti, and E. H. Bergou, “Client selection in federated learning based on gradients importance,” in AIP Conference Proceedings, vol. 3034, no. 1. AIP Publishing, 2024

    2024

  9. [17]

    Gtg-shapley: Efficient and accurate participant contribution evaluation in federated learning,

    Z. Liu, Y . Chen, H. Yu, Y . Liu, and L. Cui, “Gtg-shapley: Efficient and accurate participant contribution evaluation in federated learning,” ACM Transactions on intelligent Systems and Technology (TIST), vol. 13, no. 4, pp. 1–21, 2022

    2022

  10. [18]

    Towards efficient data valuation based on the shapley value,

    R. Jia, D. Dao, B. Wang, F. A. Hubis, N. Hynes, N. M. G ¨urel, B. Li, C. Zhang, D. Song, and C. J. Spanos, “Towards efficient data valuation based on the shapley value,” in The 22nd International Conference on Artificial Intelligence and Statistics . PMLR, 2019, pp. 1167–1176

    2019

  11. [19]

    Towards understanding biased client selection in federated learning,

    Y . J. Cho, J. Wang, and G. Joshi, “Towards understanding biased client selection in federated learning,” in International Conference on Artificial Intelligence and Statistics . PMLR, 2022, pp. 10 351–10 375

    2022

  12. [20]

    Fedfast: Going beyond average for faster training of federated recommender systems,

    K. Muhammad, Q. Wang, D. O’Reilly-Morgan, E. Tragos, B. Smyth, N. Hurley, J. Geraci, and A. Lawlor, “Fedfast: Going beyond average for faster training of federated recommender systems,” in Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & dat...

    2020

  13. [21]

    Refrs: Resource-efficient federated recommender system for dynamic and diversified user preferences,

    M. Imran, H. Yin, T. Chen, Q. V . H. Nguyen, A. Zhou, and K. Zheng, “Refrs: Resource-efficient federated recommender system for dynamic and diversified user preferences,” ACM Transactions on Information Systems, vol. 41, no. 3, pp. 1–30, 2023

    2023

  14. [22]

    Personalized federated recommendation via joint representation learning, user clustering, and model adaptation,

    S. Luo, Y . Xiao, and L. Song, “Personalized federated recommendation via joint representation learning, user clustering, and model adaptation,” in Proceedings of the 31st ACM international conference on information & knowledge management , 2022, pp. 4289–4293

    2022

  15. [23]

    Bandit-based communication-efficient client selection strategies for federated learn- ing,

    Y . J. Cho, S. Gupta, G. Joshi, and O. Ya ˘gan, “Bandit-based communication-efficient client selection strategies for federated learn- ing,” in 2020 54th Asilomar Conference on Signals, Systems, and Computers. IEEE, 2020, pp. 1066–1069

    2020

  16. [24]

    Joint client selection and bandwidth allocation algorithm for federated learning,

    H. Ko, J. Lee, S. Seo, S. Pack, and V . C. Leung, “Joint client selection and bandwidth allocation algorithm for federated learning,” IEEE Transactions on Mobile Computing, vol. 22, no. 6, pp. 3380–3390, 2021

    2021

  17. [25]

    Deep-reinforcement-learning-assisted client selection in nonorthogonal- multiple-access-based federated learning,

    R. Albelaihi, A. Alasandagutti, L. Yu, J. Yao, and X. Sun, “Deep-reinforcement-learning-assisted client selection in nonorthogonal- multiple-access-based federated learning,” IEEE Internet of Things Jour- nal, vol. 10, no. 17, pp. 15 515–15 525, 2023

    2023

  18. [26]

    A multi-agent reinforcement learning approach for efficient client selection in federated learning,

    S. Q. Zhang, J. Lin, and Q. Zhang, “A multi-agent reinforcement learning approach for efficient client selection in federated learning,” in Proceedings of the AAAI conference on artificial intelligence , vol. 36, no. 8, 2022, pp. 9091–9099

    2022

  19. [27]

    Reinforce- ment active client selection for federated heterogeneous graph learning,

    J. Wang, Y . Li, Y . Shao, Z. Xue, Z. Guan, A. Li, and G. Ye, “Reinforce- ment active client selection for federated heterogeneous graph learning,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 20, 2025, pp. 21 117–21 125

    2025

  20. [28]

    The movielens datasets: History and context,

    F. M. Harper and J. A. Konstan, “The movielens datasets: History and context,” Acm transactions on interactive intelligent systems (tiis), vol. 5, no. 4, pp. 1–19, 2015. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 14

    2015

  21. [29]

    Federated neural collaborative filter- ing,

    V . Perifanis and P. S. Efraimidis, “Federated neural collaborative filter- ing,” Knowledge-Based Systems, vol. 242, p. 108441, 2022

    2022

  22. [30]

    Fedrl: a reinforcement learning federated recommender system for efficient com- munication using reinforcement selector and hypernet generator,

    Y . Di, H. Shi, R. Ma, H. Gao, Y . Liu, and W. Wang, “Fedrl: a reinforcement learning federated recommender system for efficient com- munication using reinforcement selector and hypernet generator,” ACM Transactions on Recommender Systems , 2024

    2024

  23. [31]

    Neural col- laborative filtering,

    X. He, L. Liao, H. Zhang, L. Nie, X. Hu, and T.-S. Chua, “Neural col- laborative filtering,” in Proceedings of the 26th international conference on world wide web , 2017, pp. 173–182

    2017

  24. [32]

    Personalized item repre- sentations in federated multimodal recommendation,

    Z. Li, G. Long, J. Jiang, and C. Zhang, “Personalized item repre- sentations in federated multimodal recommendation,” arXiv preprint arXiv:2410.08478, 2024

    2024

  25. [33]

    Gpfedrec: Graph-guided personalization for federated recommendation,

    C. Zhang, G. Long, T. Zhou, Z. Zhang, P. Yan, and B. Yang, “Gpfedrec: Graph-guided personalization for federated recommendation,” in Pro- ceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , ser. KDD ’24. New York, NY , USA: Association for Compu...

    2024

  26. [34]

    When federated recommendation meets cold-start problem: Separating item attributes and user interactions,

    Zhang and et al., “When federated recommendation meets cold-start problem: Separating item attributes and user interactions,” in Proceed- ings of the ACM Web Conference 2024 , ser. WWW ’24. New York, NY , USA: Association for Computing Machinery, 2024, p. 3632–3642

    2024

  27. [35]

    Deckg: Decentralized collaborative learning with knowledge graph enhancement for poi recommendation,

    R. Zheng, L. Qu, G. Ye, T. Chen, Y . Shi, and H. Yin, “Deckg: Decentralized collaborative learning with knowledge graph enhancement for poi recommendation,” Information Sciences, p. 122570, 2025

    2025

  28. [36]

    Robust federated contrastive recommender system against targeted model poisoning attack,

    W. Yuan, C. Yang, L. Qu, G. Ye, Q. V . H. Nguyen, and H. Yin, “Robust federated contrastive recommender system against targeted model poisoning attack,” Science China Information Sciences , vol. 68, no. 4, pp. 1–16, 2025

    2025

  29. [37]

    Ptf- fsr: A parameter transmission-free federated sequential recommender system,

    W. Yuan, C. Yang, L. Qu, Q. V . Hung Nguyen, G. Ye, and H. Yin, “Ptf- fsr: A parameter transmission-free federated sequential recommender system,” ACM Transactions on Information Systems , vol. 43, no. 2, pp. 1–24, 2025

    2025

  30. [38]

    Pdc-frs: Privacy-preserving data contribution for federated recommender system,

    C. Yang, W. Yuan, L. Qu, and T. T. Nguyen, “Pdc-frs: Privacy-preserving data contribution for federated recommender system,” in International Conference on Advanced Data Mining and Applications . Springer, 2024, pp. 65–79

    2024

  31. [39]

    Robust federated contrastive recommender system against model poisoning attack,

    W. Yuan, C. Yang, L. Qu, G. Ye, Q. V . H. Nguyen, and H. Yin, “Robust federated contrastive recommender system against model poisoning attack,” arXiv preprint arXiv:2403.20107 , 2024

    2024 arXiv

  32. [40]

    Towards personalized privacy: User-governed data contribution for federated recommendation,

    L. Qu, W. Yuan, R. Zheng, L. Cui, Y . Shi, and H. Yin, “Towards personalized privacy: User-governed data contribution for federated recommendation,” in Proceedings of the ACM Web Conference 2024 , 2024, pp. 3910–3918

    2024

  33. [41]

    Semi-decentralized federated ego graph learning for recom- mendation,

    L. Qu, N. Tang, R. Zheng, Q. V . H. Nguyen, Z. Huang, Y . Shi, and H. Yin, “Semi-decentralized federated ego graph learning for recom- mendation,” in Proceedings of the ACM web conference 2023 , 2023, pp. 339–348

    2023

  34. [42]

    Hide your model: A parameter transmission-free federated recommender system,

    W. Yuan, C. Yang, L. Qu, Q. V . H. Nguyen, J. Li, and H. Yin, “Hide your model: A parameter transmission-free federated recommender system,” in 2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 2024, pp. 611–624

    2024

  35. [43]

    Hetefedrec: Federated recommender systems with model heterogeneity,

    W. Yuan, L. Qu, L. Cui, Y . Tong, X. Zhou, and H. Yin, “Hetefedrec: Federated recommender systems with model heterogeneity,” in 2024 IEEE 40th International Conference on Data Engineering (ICDE) . IEEE, 2024, pp. 1324–1337

    2024

  36. [44]

    Federated collaborative filtering for privacy-preserving personalized recommendation system,

    M. Ammad-Ud-Din, E. Ivannikova, S. A. Khan, W. Oyomno, Q. Fu, K. E. Tan, and A. Flanagan, “Federated collaborative filtering for privacy-preserving personalized recommendation system,” arXiv preprint arXiv:1901.09888, 2019

    1901 arXiv

  37. [45]

    Fedgnn: Federated graph neural network for privacy-preserving recommendation,

    C. Wu, F. Wu, Y . Cao, Y . Huang, and X. Xie, “Fedgnn: Federated graph neural network for privacy-preserving recommendation,” arXiv preprint arXiv:2102.04925, 2021

    2021 arXiv

  38. [46]

    Feddsr: Daily schedule recommendation in a federated deep reinforcement learning framework,

    W. Huang, J. Liu, T. Li, T. Huang, S. Ji, and J. Wan, “Feddsr: Daily schedule recommendation in a federated deep reinforcement learning framework,” IEEE Transactions on Knowledge and Data Engineering , vol. 35, no. 4, pp. 3912–3924, 2021

    2021

  39. [47]

    Communication-efficient learning of deep networks from decentralized data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in Artificial intelligence and statistics . PMLR, 2017, pp. 1273– 1282

    2017

  40. [48]

    Fedgst: An efficient federated graph neural network for spatio-temporal poi recommendation,

    T. Tang, M. Hou, S. Yu, Z. Cai, Z. Han, G. Oatley, and V . Saikrishna, “Fedgst: An efficient federated graph neural network for spatio-temporal poi recommendation,” ACM Transactions on Sensor Networks , 2024

    2024

  41. [49]

    Fedacs: An adaptive client selection framework for communication-efficient federated graph learning,

    H. Xu, X. Gao, J. Liu, Q. Ma, and L. Huang, “Fedacs: An adaptive client selection framework for communication-efficient federated graph learning,” IEEE Transactions on Mobile Computing , 2025

    2025

  42. [50]

    Ldp-fed: Federated learning with local differential privacy,

    S. Truex, L. Liu, K.-H. Chow, M. E. Gursoy, and W. Wei, “Ldp-fed: Federated learning with local differential privacy,” in Proceedings of the third ACM international workshop on edge systems, analytics and networking, 2020, pp. 61–66

    2020

  43. [51]

    Bpr: Bayesian personalized ranking from implicit feedback,

    S. Rendle, C. Freudenthaler, Z. Gantner, and L. Schmidt-Thieme, “Bpr: Bayesian personalized ranking from implicit feedback,” arXiv preprint arXiv:1205.2618, 2012

    2012 arXiv

  44. [52]

    Addressing function approxi- mation error in actor-critic methods,

    S. Fujimoto, H. Hoof, and D. Meger, “Addressing function approxi- mation error in actor-critic methods,” in International conference on machine learning. PMLR, 2018, pp. 1587–1596

    2018

  45. [53]

    Justifying recommendations using distantly-labeled reviews and fine-grained aspects,

    J. Ni, J. Li, and J. McAuley, “Justifying recommendations using distantly-labeled reviews and fine-grained aspects,” in Proceedings of the 2019 conference on empirical methods in natural language pro- cessing and the 9th international joint conference on natural language proce...

    2019

  46. [54]

    Secure federated matrix factorization,

    D. Chai, L. Wang, K. Chen, and Q. Yang, “Secure federated matrix factorization,” IEEE Intelligent Systems, vol. 36, no. 5, pp. 11–20, 2020

    2020

  47. [55]

    Dual personalization on federated recommendation,

    C. Zhang, G. Long, T. Zhou, P. Yan, Z. Zhang, C. Zhang, and B. Yang, “Dual personalization on federated recommendation,” in Proceedings of the Thirty-Second International Joint Conference on Artificial Intelli- gence, ser. IJCAI ’23, 2023

    2023

  48. [56]

    A federated graph neural network framework for privacy-preserving personalization,

    C. Wu, F. Wu, L. Lyu, T. Qi, Y . Huang, and X. Xie, “A federated graph neural network framework for privacy-preserving personalization,” Nature Communications, vol. 13, no. 1, p. 3091, 2022

    2022

  49. [57]

    Bars: Towards open benchmarking for recommender systems,

    J. Zhu, Q. Dai, L. Su, R. Ma, J. Liu, G. Cai, X. Xiao, and R. Zhang, “Bars: Towards open benchmarking for recommender systems,” in Pro- ceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2022, pp. 2912–2923

    2022

  50. [58]

    Understanding the difficulty of training deep feedforward neural networks,

    X. Glorot and Y . Bengio, “Understanding the difficulty of training deep feedforward neural networks,” in Proceedings of the thirteenth international conference on artificial intelligence and statistics . JMLR Workshop and Conference Proceedings, 2010, pp. 249–256

    2010

  51. [59]

    Efficient multimodal streaming recommendation via expandable side mixture-of- experts,

    Y . Qu, L. Qu, T. Chen, Q. V . H. Nguyen, and H. Yin, “Efficient multimodal streaming recommendation via expandable side mixture-of- experts,” arXiv preprint arXiv:2508.05993 , 2025

    2025

  52. [60]

    Scalable dynamic embedding size search for streaming recommendation,

    Y . Qu, L. Qu, T. Chen, X. Zhao, Q. V . H. Nguyen, and H. Yin, “Scalable dynamic embedding size search for streaming recommendation,” in Proceedings of the 33rd ACM International Conference on Information and Knowledge Management , 2024, pp. 1941–1950

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.