pith. machine review for the scientific record. sign in

arxiv: 2605.12527 · v1 · submitted 2026-04-10 · 💻 cs.IR · cs.HC

Recognition: 2 theorem links

· Lean Theorem

Beyond Centralization: User-Controlled Federated Recommendations in Practice

Manel Slokom , Alejandro Bellogin
Authors on Pith no claims yet

Pith reviewed 2026-05-14 21:39 UTC · model grok-4.3

classification 💻 cs.IR cs.HC
keywords federated learningrecommender systemsuser controlprivacy preservationpersonalizationdiversitylive deployment
0
0 comments X

The pith

Users can control the balance between personalization and diversity in a federated recommender while keeping data private.

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

The paper demonstrates a practical federated recommendation system in which users retain control over their data and can actively choose the recommendation objective. Through a live experiment lasting 53 days with 22 participants, it shows that users engage with these controls, achieve higher click-through rates with personalized rankings, and report satisfaction with the setup. This combination suggests that privacy-preserving systems need not sacrifice effectiveness if users are given direct oversight of the process. The results indicate that immediate feedback helps users understand and utilize the available controls effectively.

Core claim

In a deployment with 22 users and 8807 titles over 53 days, a federated recommender allowed participants to switch between personalization-focused and diversity-enhanced ranking objectives without sharing raw data. Users preferred the personalization option, recording 65.37% CTR compared to 62.07%, interacted with settings 248 times on average satisfaction of 3.93 out of 5, and responded to immediate feedback by adjusting their choices.

What carries the argument

The interactive federated recommendation engine that runs locally on user devices, enabling real-time switching of ranking objectives while using federated learning for model updates.

If this is right

  • Users actively engage with control interfaces when provided with immediate feedback on how changes affect recommendations.
  • Personalization objectives lead to higher engagement metrics than diversity-enhanced alternatives in user-controlled settings.
  • Privacy can be maintained through on-device data without eliminating the ability to deliver relevant recommendations.
  • Such systems foster user understanding of recommendation mechanics through direct interaction.

Where Pith is reading between the lines

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

  • Extending this to more users could reveal whether control mechanisms remain intuitive at scale.
  • Similar user-controlled approaches might apply to other domains like content moderation or advertising personalization.
  • Longer studies could test if engagement with controls persists over months rather than weeks.

Load-bearing premise

The small group of 22 participants over 53 days provides enough evidence that the system works for broader use.

What would settle it

Finding that a larger group of users shows low engagement with controls or prefers centralized systems would indicate the approach does not generalize.

Figures

Figures reproduced from arXiv: 2605.12527 by Alejandro Bellogin, Manel Slokom.

Figure 1
Figure 1. Figure 1: End-to-end demo loop. Viewing history stays on device. Local training produces model updates. We apply differential [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: User interface and control mechanisms. (a) Recommendation feed with local filtering and diversity-enhanced mode [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: CTR trend over 53 days shows stable personalization [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
read the original abstract

Recommendation systems typically require centralized user data, limiting user control and raising privacy concerns. Federated learning offers an alternative by keeping data on-device, but its impact on real user behavior remains largely unexplored. We present a live federated recommender system that allows users to control the recommendation objective while keeping their data local. In a 53-day deployment with 22 participants and a catalog of 8807 titles, users interacted with recommendations and switched between personalization and diversity-enhanced ranking. We find that users prefer personalization when given explicit choice (65.37\% vs.\ 62.07\% CTR), actively engage with control mechanisms (3.93/5 satisfaction; 248 settings changes), and develop an understanding of how their interactions affect recommendations through immediate feedback. Our results show that user control, privacy, and effective personalization can be combined in a working system. We demonstrate a practical approach to interactive, privacy-preserving recommendation. Code and demo materials are available at: https://github.com/SlokomManel/federated-recommendations-participants

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

1 major / 1 minor

Summary. The paper reports on the design and a 53-day live deployment of a federated recommender system with 22 participants and an 8807-title catalog. Users can interactively switch between personalization and diversity-enhanced ranking objectives while keeping data on-device. Results include a CTR of 65.37% for personalization versus 62.07% for diversity, mean satisfaction of 3.93/5, and 248 control-setting changes, leading to the claim that user control, privacy, and effective personalization can be combined in a working practical system.

Significance. If the findings generalize, the work supplies rare real-user empirical data on interactive federated recommendation, showing that on-device control mechanisms can be used without sacrificing engagement. The public release of code and demo materials at the cited GitHub repository is a clear strength for reproducibility.

major comments (1)
  1. [Deployment and Results] Deployment and Results section: The central claim that the system constitutes a 'practical approach' and demonstrates that 'user control, privacy, and effective personalization can be combined' rests on N=22 participants over 53 days. No participant-recruitment details, selection-bias discussion, statistical-power analysis, or significance tests for the CTR difference (65.37% vs 62.07%) are reported, so the evidence for general viability remains limited.
minor comments (1)
  1. [Abstract] The abstract states that users 'develop an understanding of how their interactions affect recommendations through immediate feedback,' but the main text provides no concrete metric or example illustrating this understanding.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback. We agree that the small-scale live deployment requires clearer contextualization of its limitations and will revise the manuscript to add the requested details on recruitment, bias, power, and statistical testing while preserving the value of the real-world interactive data.

read point-by-point responses

  1. Referee: [Deployment and Results] Deployment and Results section: The central claim that the system constitutes a 'practical approach' and demonstrates that 'user control, privacy, and effective personalization can be combined' rests on N=22 participants over 53 days. No participant-recruitment details, selection-bias discussion, statistical-power analysis, or significance tests for the CTR difference (65.37% vs 62.07%) are reported, so the evidence for general viability remains limited.

    Authors: We agree the evidence for broad general viability is limited by the sample size. In revision we will add: participant recruitment details (voluntary opt-in from a university community via privacy-oriented channels with informed consent); explicit discussion of selection bias (likely over-representation of privacy-conscious or technically interested users); a post-hoc power analysis noting modest power for small effects; and a two-proportion z-test for the CTR difference with the resulting p-value. We will also insert a dedicated Limitations subsection that tempers language around 'practical approach' to emphasize the study's exploratory character while underscoring the novelty of on-device interactive control. These additions will be placed in the Deployment and Results section. revision: yes

Circularity Check

0 steps flagged

No significant circularity: empirical deployment study with direct observations

full rationale

The paper reports results from a live 53-day deployment involving 22 participants interacting with a federated recommender system, including CTR comparisons (65.37% vs 62.07%), satisfaction ratings (3.93/5), and counts of setting changes (248). No mathematical derivations, equations, fitted parameters, or predictive models are described that could reduce to inputs by construction. The central claim—that user control, privacy, and personalization can coexist in a working system—is supported solely by these empirical metrics from the deployment itself, with no self-citations, ansatzes, or uniqueness theorems invoked as load-bearing elements. This is a standard empirical evaluation without any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on empirical observations from a user study rather than mathematical derivations or new theoretical constructs.

axioms (1)
  • domain assumption User satisfaction and click-through rates accurately reflect preference for recommendation objectives.
    The paper interprets CTR and survey scores as indicators of preference without additional validation.

pith-pipeline@v0.9.0 · 5476 in / 1048 out tokens · 72538 ms · 2026-05-14T21:39:44.068562+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

50 extracted references · 50 canonical work pages

  1. [1]

    and Runeson, P

    Wohlin, C. and Runeson, P. and. 2012 , Keywords =

    work page 2012

  2. [2]

    and Wong, Kok-Seng , year=

    Nguyen, Ngoc-Hieu and Nguyen, Tuan-Anh and Nguyen, Tuan and Hoang, Vu Tien and Le, Dung D. and Wong, Kok-Seng , year=. Towards Efficient Communication and Secure Federated Recommendation System via Low-rank Training , url=. doi:10.1145/3589334.3645702 , booktitle=

    work page doi:10.1145/3589334.3645702

  3. [3]

    2021 , isbn =

    Anelli, Vito Walter and Deldjoo, Yashar and Di Noia, Tommaso and Ferrara, Antonio and Narducci, Fedelucio , title =. 2021 , isbn =. doi:10.1145/3412841.3442010 , booktitle =

    work page doi:10.1145/3412841.3442010 2021

  4. [4]

    2012 , isbn =

    Pariser, Eli , title =. 2012 , isbn =

    work page 2012

  5. [5]

    Federated Recommendation Systems , year=

    Qiang, Yang , booktitle=. Federated Recommendation Systems , year=

    work page

  6. [6]

    Differential Privacy: A Survey of Results , booktitle=

    Dwork, Cynthia , editor=. Differential Privacy: A Survey of Results , booktitle=. 2008 , publisher=

    work page 2008

  7. [7]

    Late breaking results, in conjunction with the 13th ACM International Conference on Recommender Systems , year=

    Data Masking for Recommender Systems: Prediction Performance and Rating Hiding , author=. Late breaking results, in conjunction with the 13th ACM International Conference on Recommender Systems , year=

    work page

  8. [8]

    Information Processing & Management , volume =

    Towards user-oriented privacy for recommender system data: A personalization-based approach to gender obfuscation for user profiles , author =. Information Processing & Management , volume =. 2021 , issn =

    work page 2021

  9. [9]

    FDRS: Federated Diversified Recommender System Based on Heterogeneous Graph Convolutional Network , year=

    Li, Nianzhe and Liu, Hanwen and Meng, Shunmei and Li, Qianmu , booktitle=. FDRS: Federated Diversified Recommender System Based on Heterogeneous Graph Convolutional Network , year=

    work page

  10. [10]

    FedRec: Federated Recommendation With Explicit Feedback , year=

    Lin, Guanyu and Liang, Feng and Pan, Weike and Ming, Zhong , journal=. FedRec: Federated Recommendation With Explicit Feedback , year=

    work page

  11. [11]

    2019 , eprint=

    Federated Collaborative Filtering for Privacy-Preserving Personalized Recommendation System , author=. 2019 , eprint=

    work page 2019

  12. [12]

    McMahan, James and Ramage, Daniel , year =

    work page

  13. [13]

    The Use of MMR, Diversity-Based Reranking for Reordering Documents and Producing Summaries , journal =

    Carbonell, Jaime and Stewart, Jade , year =. The Use of MMR, Diversity-Based Reranking for Reordering Documents and Producing Summaries , journal =

    work page

  14. [14]

    PySyft: A Library for Easy Federated Learning , isbn =

    Ziller, Alexander and Trask, Andrew and Lopardo, Antonio and Szymkow, Benjamin and Wagner, Bobby and Bluemke, Emma and Nounahon, Jean-Mickael and Passerat-Palmbach, Jonathan and Prakash, Kritika and Rose, Nick and Ryffel, Théo and Reza, Zarreen Naowal and Kaissis, Georgios , year =. PySyft: A Library for Easy Federated Learning , isbn =

    work page

  15. [15]

    Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence , pages =

    Rendle, Steffen and Freudenthaler, Christoph and Gantner, Zeno and Schmidt-Thieme, Lars , title =. Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence , pages =. 2009 , isbn =

    work page 2009

  16. [16]

    Lauw , title =

    Aghiles Salah and Quoc-Tuan Truong and Hady W. Lauw , title =. Journal of Machine Learning Research , year =

    work page

  17. [18]

    1994 , url=

    The Goal Question Metric Approach , author=. 1994 , url=

    work page 1994

  18. [19]

    Multi-objective optimization with recommender systems: A systematic review , journal =

    Fatima Ezzahra Zaizi and Sara Qassimi and Said Rakrak , keywords =. Multi-objective optimization with recommender systems: A systematic review , journal =. 2023 , issn =. doi:https://doi.org/10.1016/j.is.2023.102233 , url =

    work page doi:10.1016/j.is.2023.102233 2023

  19. [20]

    A survey on multi-objective recommender systems , volume =

    Jannach, Dietmar and Abdollahpouri, Himan , year =. A survey on multi-objective recommender systems , volume =. Frontiers in big data , doi =

    work page

  20. [21]

    2024 , isbn =

    Tommasel, Antonela , title =. 2024 , isbn =. doi:10.1145/3640457.3688182 , booktitle =

    work page doi:10.1145/3640457.3688182 2024

  21. [22]

    and Khaokaew, Yonchanok and Chan, Jeffrey , title =

    Cruz, Francis Zac Dela and Salim, Flora D. and Khaokaew, Yonchanok and Chan, Jeffrey , title =. 2024 , isbn =. doi:10.1145/3640457.3688170 , booktitle =

    work page doi:10.1145/3640457.3688170 2024

  22. [23]

    and Derr, Tyler , title =

    Zhao, Yuying and Wang, Yu and Liu, Yunchao and Cheng, Xueqi and Aggarwal, Charu C. and Derr, Tyler , title =. 2025 , issue_date =. doi:10.1145/3664928 , journal =

    work page doi:10.1145/3664928 2025

  23. [24]

    2009 , isbn =

    Zhang, Mi , title =. 2009 , isbn =. doi:10.1145/1639714.1639798 , booktitle =

    work page doi:10.1145/1639714.1639798 2009

  24. [25]

    2024 , isbn =

    Eleftherakis, Stavroula and Koutrika, Georgia and Amer-Yahia, Sihem , title =. 2024 , isbn =. doi:10.1145/3627043.3659539 , booktitle =

    work page doi:10.1145/3627043.3659539 2024

  25. [26]

    2019 , isbn =

    Sacharidis, Dimitris , title =. 2019 , isbn =. doi:10.1145/3297280.3297442 , booktitle =

    work page doi:10.1145/3297280.3297442 2019

  26. [27]

    2020 , isbn =

    Mansoury, Masoud and Abdollahpouri, Himan and Pechenizkiy, Mykola and Mobasher, Bamshad and Burke, Robin , title =. 2020 , isbn =. doi:10.1145/3340631.3394860 , booktitle =

    work page doi:10.1145/3340631.3394860 2020

  27. [28]

    2019 , isbn =

    Zou, Lixin and Xia, Long and Ding, Zhuoye and Yin, Dawei and Song, Jiaxing and Liu, Weidong , title =. 2019 , isbn =. doi:10.1007/978-3-030-18579-4_7 , booktitle =

    work page doi:10.1007/978-3-030-18579-4_7 2019

  28. [29]

    COSPLAY: Concept Set Guided Personalized Dialogue Generation Across Both Party Personas , url=

    Naghiaei, Mohammadmehdi and Rahmani, Hossein A. and Deldjoo, Yashar , title =. 2022 , isbn =. doi:10.1145/3477495.3531959 , booktitle =

    work page doi:10.1145/3477495.3531959 2022

  29. [30]

    How to Diversify any Personalized Recommender?

    Slokom, Manel and Daniil, Savvina and Hollink, Laura. How to Diversify any Personalized Recommender?. Advances in Information Retrieval. 2025

    work page 2025

  30. [31]

    InProceedings of the ACM Web Conference 2024 (WWW ’24)

    Wang, Yifan and Sun, Peijie and Ma, Weizhi and Zhang, Min and Zhang, Yuan and Jiang, Peng and Ma, Shaoping , title =. 2024 , isbn =. doi:10.1145/3589334.3645518 , booktitle =

    work page doi:10.1145/3589334.3645518 2024

  31. [32]

    and Yan, Shen and He, Chaoyang and Ferrara, Emilio and Avestimehr, Salman , title =

    Ezzeldin, Yahya H. and Yan, Shen and He, Chaoyang and Ferrara, Emilio and Avestimehr, Salman , title =. 2023 , isbn =. doi:10.1609/aaai.v37i6.25911 , booktitle =

    work page doi:10.1609/aaai.v37i6.25911 2023

  32. [33]

    2019 , issue_date =

    Xia, Bin and Yin, Junjie and Xu, Jian and Li, Yun , title =. 2019 , issue_date =. doi:10.1016/j.knosys.2019.07.028 , journal =

    work page doi:10.1016/j.knosys.2019.07.028 2019

  33. [34]

    2022 , issue_date =

    Wu, Haolun and Ma, Chen and Mitra, Bhaskar and Diaz, Fernando and Liu, Xue , title =. 2022 , issue_date =. doi:10.1145/3564285 , journal =

    work page doi:10.1145/3564285 2022

  34. [35]

    A hybrid recommendation system with many-objective evolutionary algorithm , journal =

    Xingjuan Cai and Zhaoming Hu and Peng Zhao and WenSheng Zhang and Jinjun Chen , keywords =. A hybrid recommendation system with many-objective evolutionary algorithm , journal =. 2020 , issn =. doi:https://doi.org/10.1016/j.eswa.2020.113648 , url =

    work page doi:10.1016/j.eswa.2020.113648 2020

  35. [36]

    2024 , isbn =

    Boratto, Ludovico and Fabbri, Francesco and Fenu, Gianni and Marras, Mirko and Medda, Giacomo , title =. 2024 , isbn =. doi:10.1145/3640457.3688064 , booktitle =

    work page doi:10.1145/3640457.3688064 2024

  36. [37]

    A Survey of Personalized News Recommendation , volume =

    Meng, Xiangfu and Huo, Hongjin and Zhang, Xiaoyan and Wang, Wanchun and Zhu, Jinxia , year =. A Survey of Personalized News Recommendation , volume =. Data Science and Engineering , doi =

    work page

  37. [38]

    2011 , isbn =

    Pariser, Eli , title =. 2011 , isbn =

    work page 2011

  38. [39]

    Novelty and Diversity in Recommender Systems

    Castells, Pablo and Hurley, Neil and Vargas, Sa \'u l. Novelty and Diversity in Recommender Systems. Recommender Systems Handbook. 2022. doi:10.1007/978-1-0716-2197-4_16

    work page doi:10.1007/978-1-0716-2197-4_16 2022

  39. [40]

    RADio – Rank-Aware Divergence Metrics to Measure Normative Diversity in News Recommendations , year =

    Vrijenhoek, Sanne and B\'. RADio – Rank-Aware Divergence Metrics to Measure Normative Diversity in News Recommendations , year =. doi:10.1145/3523227.3546780 , booktitle =

    work page doi:10.1145/3523227.3546780

  40. [41]

    Recommender Systems Handbook

    work page

  41. [42]

    ResearchGate , year =

    work page

  42. [43]

    2011 , isbn =

    Bellogin, Alejandro and Castells, Pablo and Cantador, Ivan , title =. 2011 , isbn =. doi:10.1145/2043932.2043996 , booktitle =

    work page doi:10.1145/2043932.2043996 2011

  43. [44]

    A Survey on Federated Recommendation Systems , year=

    Sun, Zehua and Xu, Yonghui and Liu, Yong and He, Wei and Kong, Lanju and Wu, Fangzhao and Jiang, Yali and Cui, Lizhen , journal=. A Survey on Federated Recommendation Systems , year=

    work page

  44. [45]

    , title =

    Dong, Jinshuo and Roth, Aaron and Su, Weijie J. , title =. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume =. 2022 , issn =. doi:10.1111/rssb.12454 , url =

    work page doi:10.1111/rssb.12454 2022

  45. [46]

    2015 , eprint=

    Optimality of the Laplace Mechanism in Differential Privacy , author=. 2015 , eprint=

    work page 2015

  46. [47]

    2013 , isbn =

    Steck, Harald , title =. 2013 , isbn =. doi:10.1145/2507157.2507160 , booktitle =

    work page doi:10.1145/2507157.2507160 2013

  47. [48]

    Rating prediction of recommended item based on review deep learning and rating probability matrix factorization , journal =

    Zhiguo Zhu and Mengru Yan and Xiaoyi Deng and Ming Gao , keywords =. Rating prediction of recommended item based on review deep learning and rating probability matrix factorization , journal =. 2022 , issn =. doi:https://doi.org/10.1016/j.elerap.2022.101160 , url =

    work page doi:10.1016/j.elerap.2022.101160 2022

  48. [49]

    Top-N-Rank: A Scalable List-wise Ranking Method for Recommender Systems , year=

    Liang, Junjie and Hu, Jinlong and Dong, Shoubin and Honavar, Vasant , booktitle=. Top-N-Rank: A Scalable List-wise Ranking Method for Recommender Systems , year=

    work page

  49. [50]

    Proceedings of the 26th Annual Conference on Learning Theory , pages =

    A Theoretical Analysis of NDCG Type Ranking Measures , author =. Proceedings of the 26th Annual Conference on Learning Theory , pages =. 2013 , editor =

    work page 2013

  50. [51]

    How good your recommender system is? A survey on evaluations in recommendation , volume =

    Silveira, Thiago and Zhang, Min and Lin, Xiao and Liu, Yiqun and Ma, Shaoping , year =. How good your recommender system is? A survey on evaluations in recommendation , volume =. International Journal of Machine Learning and Cybernetics , doi =

    work page