Pith. sign in

REVIEW 3 major objections 5 minor 44 references

Personalized Language Model Learning on Text Data Without User Identifiers

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Cloud language models can be personalized from identifier-free text by sampling each user's embedding fresh from a per-user distribution; the paper proves anonymization requires distributions to be non-identifiable (e.g., Beta) or close…

desk verdict A genuinely new personalization mechanism with a plausible accuracy story, but the privacy guarantee is argued for the wrong object: the cloud sees text and repeated queries, not just a mixture of embeddings. read the letter →

arxiv 2501.06062 v1 pith:MU4U6WE3 submitted 2025-01-10 cs.LG

Yucheng Ding , Yangwenjian Tan , Xiangyu Liu , Chaoyue Niu , Fandong Meng , Jie Zhou , Ning Liu , Fan Wu
show 1 more author
Guihai Chen
This is my paper · ORCID
classification cs.LG
keywords personalizedlanguagemodelsanonymousdatauploaduserembeddingnon-identifiabilitymisattributionon-devicetrainingcloudinferenceprompttuning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes that a cloud language model can be personalized over text uploaded without user identifiers if each device maintains a user-specific distribution for user embeddings and samples a fresh embedding for every query. That breaks the one-to-one link between a fixed embedding and a person, so the cloud sees a mixture rather than identity tags. The authors prove two ways to keep the cloud from tracing an embedding to its source: either the distribution family is linearly dependent, making the mixture non-identifiable (Beta distributions are the worked example), or the per-user distributions are kept close enough that attributing an embedding to the right user fails with high probability. On three public datasets and one industrial mobile-keyboard dataset, the proposed framework improves inference accuracy by up to 5.69 percent over cloud training without personalization while adding only about 0.01 seconds of latency.

What carries the argument

The machinery is the user-specific embedding distribution $U_n$, a probability law on the embedding space maintained on the user's device, replacing the static user embedding. For the non-identifiable route, the family of CDFs is chosen linearly dependent; the Beta family works because any Beta density satisfies $f(\alpha,\beta) = c_1 f(\alpha+1,\beta) + c_2 f(\alpha,\beta+1)$, so the observed mixture admits multiple decompositions. For the misattribution route, users start from the same initialization and train with a bounded learning rate, gradient clipping, and a shared variance $\sigma^2$, which keeps the means within $2\|\eta T G\|$; the theorem then lower-bounds the probability that a sampled embedding is assigned to the wrong user. Reparameterization through an auxiliary noise variable makes gradient-based training of the distribution parameters possible while the language model is frozen.

What would settle it

On the released datasets, run an authorship-attribution or stylometric classifier over the text portions of the uploaded samples; if it can group samples by user significantly better than chance, the deployed protocol leaks identity even though the embedding-attribution theorems say a single embedding is ambiguous.

Watch

Extended reading notes

Core claim

The central claim is that the contradiction between anonymization and personalization disappears once a user is represented not by a fixed embedding but by a distribution. Each device trains the parameters of its own embedding distribution against the frozen language model using the reparameterization trick, then uploads text paired with a freshly sampled embedding, and the cloud fine-tunes on the resulting mixture. Tracking is prevented either because the mixture is non-identifiable — for Beta distributions, each component density is itself a combination of two other Beta densities, so infinitely many decompositions exist — or because per-user Gaussians are kept so close that the best-attribution rule is wrong with probability at least $1 - \left(\Phi(\|\eta T G\|/\sigma)\right)^{N-1}$. Experiments across three public datasets and one industrial task show accuracy gains up to 5.69 percent over cloud training without personalization, at roughly 0.01 seconds added latency.

Load-bearing premise

The privacy guarantees collapse if the cloud can connect repeated uploads from the same person, or can use the text that arrives beside each sampled embedding to identify its source, because the protocol's anonymity analysis treats each embedding as an isolated sample from a known mixture.

Editorial extensions

If this is right

  • Cloud-trained language models can deliver personalized predictions on text uploaded without any user identifier, improving inference accuracy by up to 5.69 percent over the no-personalization baseline.
  • The added inference cost is small enough for real-time use: about 0.01 seconds on public datasets and at most 3 milliseconds on the industrial tasks, versus hundreds of milliseconds to tens of seconds for on-device personalization.
  • Anonymization is a tunable design choice: choosing a linearly dependent distribution family (e.g., Beta) or keeping Gaussian distributions close via small learning rates, gradient clipping, and larger variance controls the trade-off between privacy and accuracy.
  • The accuracy gain is concentrated on samples whose predictions are sensitive to which user embedding is present, so personalization helps exactly the data points that vary across users.
  • The same recipe transfers across decoder-only and encoder-decoder transformer models and across classification, next-word generation, pinyin input, and word-association tasks.

Reading between the lines

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

  • One implication left implicit is that the cloud's view in the deployed protocol is richer than the theorems model: each uploaded sample pairs an embedding with the user's text, and a single user uploads many samples over time, so text content and temporal patterns could be used to link samples even if the embeddings alone are ambiguous.
  • If the mechanism is adopted, the variance and learning-rate settings from the misattribution theorem give operators a practical privacy knob: they can choose a target misattribution probability and accept the measured accuracy loss, as the paper's variance sweep illustrates.
  • The same distribution-sampling trick should transfer to other cloud services that consume user embeddings, such as recommendation or image models, whenever a per-user distribution can be trained on-device and the embedding dimension is small enough to upload cheaply.
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 proposes IDfree-PL, a framework in which each mobile device learns a user-specific distribution over user embeddings (instantiated with Beta or Gaussian distributions) and uploads a freshly sampled embedding concatenated with the raw text to the cloud; the cloud then fine-tunes a language model on these tuples and serves real-time inference. The authors prove non-identifiability of Beta mixtures and a lower bound on the misattribution probability of Gaussian embeddings, and they report accuracy gains of up to 5.69% over a no-personalization baseline with negligible added latency across three public datasets and one industrial dataset. The central advertised contribution is that the cloud can be prevented from tracking users through uploaded embeddings while still obtaining personalized service.

Significance. If the privacy guarantee were established, this would be a practically relevant contribution: it addresses a real deployment constraint for cloud-based personalized language models, and the experimental apparatus is substantial, including three public datasets, one industrial dataset, four models, and a released code repository. The mathematical derivations in Theorems 1 and 2 are genuine and the Beta-identifiability argument is correct as a statement about mixture identifiability. The main weakness is that the theorems are proved for the marginal distribution of the sampled embedding only, whereas the protocol exposes the user's raw text alongside every embedding; the paper's broad privacy claims therefore go beyond what the analysis supports.

major comments (3)
  1. [Algorithm 1 / Section 3.1 and 4.2] The anonymization theorems analyze only the marginal distribution of the sampled user embedding u_n, but the protocol uploads ([u_n; x], y) in training (line 6) and [u_n; x] in inference (line 9). The cloud therefore observes the raw text x together with every embedding. X is drawn from a user-specific language distribution that is not part of the Beta mixture in Theorem 1 or the Gaussian model in Theorem 2. Non-identifiability of the marginal mixture over u_n does not imply non-identifiability of the joint distribution over (u_n, x), and an adversary who conditions on x, or on the pair (u_n, x), can attribute samples and link sessions far more accurately than the embedding-only analysis suggests. The central claim that the cloud cannot track users is therefore not supported by the stated theorems.
  2. [Appendix A.2 / Eq. (14)] Assumption 1 asserts that the events Pr(u_n | U_i) ≤ Pr(u_n | U_n) for different i are independent for a fixed u_n. These events are functions of the same random vector u_n; for Gaussian components with common covariance, they are half-space events with respect to u_n and are generally positively correlated. For positively correlated events, P(∩_i A_i) can exceed ∏_i P(A_i), so the inequality P(∀k≠n: ||u_n−μ_n||≤||u_n−μ_k||) ≤ (Φ(||ηTG||/σ))^{N−1} is not established and may fail in the direction needed. Consequently the lower bound on the misattribution probability in Eq. (7) is not valid as proven.
  3. [Section 4.2 / Theorem 2] Even if Assumption 1 were granted, Theorem 2 concerns a single embedding drawn from a known mixture with equal priors. In the actual protocol, each user uploads many samples over time, and the cloud accumulates a dataset with timestamps and text content. Aggregation over multiple samples, together with textual style and timing signals, can be used to cluster or link a user's uploads even when each individual embedding has a high misattribution probability. The paper provides no unlinkability or longitudinal-tracking analysis, so the 'many-to-one mapping' is not established as a practical privacy guarantee.
minor comments (5)
  1. [Objective 2 / Eq. (3)] In the second summation of Eq. (3), the distribution is written as F(U_n; θ'_m); it should be F(U_m; θ'_m) (or F(U; θ'_m)) to make the alternative representation well-defined.
  2. [Table 2] The reported 'misattribution probability' is estimated by comparing sampled embeddings with distribution means rather than by evaluating the posterior in Eq. (4); this empirical proxy should be described and justified, since it is not the same quantity as the theoretical bound.
  3. [Section 5.1] The implementation description states 'we set the max norm of gradient clipping to 5' without specifying the norm type, and the relationship between this setting and the variance values in Table 2 should be clarified.
  4. [Algorithm 1 / lines 6 and 9] The word 'anonymously' in lines 6 and 9 conflates the absence of a user identifier with a privacy guarantee; I suggest replacing it with 'without a user identifier' to avoid overclaiming.
  5. [Figure 3] The speedup factors are reported without error bars or statistical significance tests; adding variance estimates across runs would make the latency comparison more informative.

Circularity Check

1 steps flagged · score 4.0 of 10

Privacy guarantee is definitional: 'anonymization of user embedding' is defined as exactly the properties Theorems 1–2 prove, so the conclusion of 'data anonymization' is a restatement of the paper's own objective rather than an independent derivation.

  1. self definitional [Section 2.3 (Objectives 2 and 3); abstract conclusion 'keeping data anonymization'; Section 8 conclusion]
    "We thus define the anonymization of user embedding from either (1) non-identifiability of user-specific distributions from the mixture of distribution in objective 2, intuitively forming one-way user-embedding mapping; or (2) wrongly attributing collected user embeddings to a specific user's distribution, even when all user-specific distributions are known, in objective 3, intuitively forming many-to-one user-embedding mapping."

    Theorem 1 and Theorem 2 prove exactly the properties that Section 2.3 stipulates as 'anonymization of user embedding': non-identifiability of the mixture (Objective 2) and high misattribution probability (Objective 3). The abstract and Section 8 then assert 'keeping data anonymization' and 'guarantee data anonymization' as if these theorems established an independent privacy result. But the guarantee is the definition: the informal target 'prevent the cloud from tracking users' is made true by defining anonymization as the two embedding-mixture properties that are subsequently proved. The step 'Theorem 1/2 => data anonymization' therefore reduces by construction to the paper's own definition.

full rationale

The mathematical derivation chain is mostly non-circular. Lemma 1 is a classical external result (Yakowitz and Spragins 1968), not a self-citation. Theorem 1's Beta identity genuinely establishes linear dependence of the Beta family and hence non-identifiability; Theorem 2's Gaussian bound follows from the stated gradient-clipping assumptions via a real geometric argument. The personalization objective and the accuracy/latency evaluations are also empirically grounded. The one significant circularity is the privacy conclusion: 'anonymization of user embedding' is explicitly defined as the two properties that Theorems 1-2 prove, and the paper then presents those theorems as demonstrating data anonymization. This is a definitional equivalence, not a derived consequence. In addition, the protocol uploads the user's raw text with every embedding and supports repeated queries, so the threat model in the theorems (a single embedding's marginal distribution) does not cover the information actually available to the cloud. That gap is a scope limitation but also part of why the claimed guarantee is forced by the paper's own definition rather than by an independent attack model. Given that the core statistical theorems and the personalization results retain independent content, the circularity is partial, not total.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical or ontological entities are introduced; the user-specific embedding distribution is a methodological construct rather than an independently testable entity. The free parameters are the variance, learning rate, and local update bounds that appear in the theoretical privacy bound, plus the learned distribution parameters themselves.

free parameters (3)
  • Gaussian embedding variance sigma^2 = 0.2 default; swept 0.02 to 0.32
    Appears directly in Theorem 2's misattribution bound and is tuned by hand to trade privacy against personalization.
  • On-device learning rate eta = 1e-3
    Chosen small to keep user-specific means close together, which increases the misattribution bound in Theorem 2.
  • Local update budget T and gradient clip norm G = T corresponds to 15 epochs; max norm 5
    Assumptions 2 and 3 bound local deviations; these values are chosen by hand and control the privacy-utility tradeoff.
assumptions (4)
  • standard math Mixture identifiability of a family F is equivalent to linear independence of F (Lemma 1, from Yakowitz and Spragins 1968)
    Used to justify the Beta-based non-identifiability result in Theorem 1; the proof is deferred to a cited theorem.
  • domain assumption The cloud cannot link multiple uploads from the same user and cannot use text content for re-identification
    Required for Objectives 2 and 3 to be interpreted as data anonymization. Algorithm 1 uploads text next to each embedding, so this assumption is hidden and load-bearing.
  • domain assumption Equal prior Pr(U_k) for all users in Bayesian attribution
    Theorem 2 states its bound only when user priors are equal; a cloud with knowledge of user activity levels would have informative priors and lower misattribution.
  • domain assumption Assumption 1: pairwise misattribution events are independent; Assumption 2: bounded local iterations T; Assumption 3: bounded clipped gradient norm
    These assumptions are needed for the closed-form misattribution bound in Theorem 2; the independence assumption is strong and not validated empirically.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Personalized Language Model Learning on Text Data Without User Identifiers." pith.science (2026). https://pith.science/paper/MU4U6WE3

@misc{pith2026250106062,
  author       = {Pith},
  title        = {Pith review of: Personalized Language Model Learning on Text Data Without User Identifiers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MU4U6WE3}},
  note         = {Machine review of arXiv:2501.06062}
}
read the original abstract

In many practical natural language applications, user data are highly sensitive, requiring anonymous uploads of text data from mobile devices to the cloud without user identifiers. However, the absence of user identifiers restricts the ability of cloud-based language models to provide personalized services, which are essential for catering to diverse user needs. The trivial method of replacing an explicit user identifier with a static user embedding as model input still compromises data anonymization. In this work, we propose to let each mobile device maintain a user-specific distribution to dynamically generate user embeddings, thereby breaking the one-to-one mapping between an embedding and a specific user. We further theoretically demonstrate that to prevent the cloud from tracking users via uploaded embeddings, the local distributions of different users should either be derived from a linearly dependent space to avoid identifiability or be close to each other to prevent accurate attribution. Evaluation on both public and industrial datasets using different language models reveals a remarkable improvement in accuracy from incorporating anonymous user embeddings, while preserving real-time inference requirement.

Figures

Figures reproduced from arXiv: 2501.06062 by the authors.

Figure 1
Figure 1. Workflow of IDfree-PL. In the training phase, (a) each mobile device first trains an optimal user distribution [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Rating classification accuracy over Amazon-Kindle [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. IDfree-PL vs. Baselines from inference latency with different language models over different datasets. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Inference accuracy of test samples with varying sensitivity to user embeddings over Amazon-Kindle dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Visualization of user embeddings. consistent with Theorem 2. More specifically, when the variance is set to 0.0 2 , which indicates that the user embeddings are static, IDfree-PL achieves the averaged inference accuracy of 80.0% but the misattribution probability is 0.…
Figure 6
Figure 6. Figure 6: T-SNE visualization of user embeddings under Gaussian and Beta distributions with different models over Amazon [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 29 canonical work pages

  1. [1]

    Apple. 2017. Deep Learning for Siri’s Voice: On-device Deep Mixture Density Networks for Hybrid Unit Selection Synthesis. https://machinelearning.apple. com/research/siri-voices

    work page 2017

  2. [2]

    Heena Reyaz Bhat, Tanveer Ahmad Lone, and Zubair M Paul. 2017. Cortana- intelligent personal digital assistant: a review. International Journal of Advanced Research in Computer Science 8, 7 (2017), 55–57

    work page 2017

  3. [3]

    Brendan McMahan, Virginia Smith, and Ameet Talwalkar

    Sebastian Caldas, Peter Wu, Tian Li, Jakub Konečný, H. Brendan McMahan, Virginia Smith, and Ameet Talwalkar. 2018. LEAF: A Benchmark for Federated Settings. CoRR abs/1812.01097 (2018)

    arXiv 2018

  4. [4]

    Zhangming Chan, Juntao Li, Xiaopeng Yang, Xiuying Chen, Wenpeng Hu, Dongyan Zhao, and Rui Yan. 2019. Modeling Personalization in Continuous Space for Response Generation via Augmented Wasserstein Autoencoders. In EMNLP/IJCNLP (1). Association for Computational Linguistics, 1931–1940

    work page 2019

  5. [5]

    Bo Chen, Yichao Wang, Zhirong Liu, Ruiming Tang, Wei Guo, Hongkun Zheng, Weiwei Yao, Muyu Zhang, and Xiuqiang He. 2021. Enhancing Explicit and Implicit Feature Interactions via Information Sharing for Parallel Deep CTR Models. In CIKM. ACM, 3757–3766

    work page 2021

  6. [6]

    Mingqing Chen, Ananda Theertha Suresh, Rajiv Mathews, Adeline Wong, Cyril Allauzen, Françoise Beaufays, and Michael Riley. 2019. Federated Learning of N-Gram Language Models. In CoNLL. Association for Computational Linguistics, 121–130

    work page 2019

  7. [7]

    Hao Cheng, Shuo Wang, Wensheng Lu, Wei Zhang, Mingyang Zhou, Kezhong Lu, and Hao Liao. 2023. Explainable Recommendation with Personalized Re- view Retrieval and Aspect Learning. In ACL (1). Association for Computational Linguistics, 51–64

    work page 2023

  8. [8]

    Liam Collins, Hamed Hassani, Aryan Mokhtari, and Sanjay Shakkottai. 2021. Exploiting Shared Representations for Personalized Federated Learning. In ICML (Proceedings of Machine Learning Research, Vol. 139) . PMLR, 2089–2099

    work page 2021

Show all 44 references
  1. [9]

    Yuyang Deng, Mohammad Mahdi Kamani, and Mehrdad Mahdavi. 2020. Adaptive Personalized Federated Learning. CoRR abs/2003.13461 (2020)

    2020 arXiv

  2. [10]

    Keyu Ding, Yongcan Wang, Zihang Xu, Zhenzhen Jia, Shijin Wang, Cong Liu, and Enhong Chen. 2023. Generative Input: Towards Next-Generation Input Methods Paradigm. CoRR abs/2311.01166 (2023)

    2023 arXiv

  3. [11]

    Ozdaglar

    Alireza Fallah, Aryan Mokhtari, and Asuman E. Ozdaglar. 2020. Personalized Fed- erated Learning with Theoretical Guarantees: A Model-Agnostic Meta-Learning Approach. In NeurIPS

    2020

  4. [12]

    Alec Go, Richa Bhayani, and Lei Huang. 2009. Twitter sentiment classification using distant supervision. CS224N project report, Stanford 1, 12 (2009), 2009

    2009

  5. [13]

    Google. 2024. https://assistant.google.com/. https://assistant.google.com/

    2024

  6. [14]

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: A Factorization-Machine based Neural Network for CTR Prediction. In IJCAI. ijcai.org, 1725–1731

    2017

  7. [15]

    Andrew Hard, Kanishka Rao, Rajiv Mathews, Françoise Beaufays, Sean Augen- stein, Hubert Eichner, Chloé Kiddon, and Daniel Ramage. 2018. Federated Learn- ing for Mobile Keyboard Prediction. CoRR abs/1811.03604 (2018)

    2018 arXiv

  8. [16]

    Yupeng Hou, Jiacheng Li, Zhankui He, An Yan, Xiusi Chen, and Julian J. McAuley

  9. [17]

    Reddi, Se- bastian U

    Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J. Reddi, Se- bastian U. Stich, and Ananda Theertha Suresh. 2020. SCAFFOLD: Stochastic Controlled Averaging for Federated Learning. In ICML (Proceedings of Machine Learning Research, Vol. 119). PMLR, 5132–5143

    2020

  10. [18]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. 2014. Auto-Encoding Variational Bayes. In ICLR

    2014

  11. [19]

    Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The Power of Scale for Parameter-Efficient Prompt Tuning. InEMNLP (1). Association for Computational Linguistics, 3045–3059

    2021

  12. [20]

    Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Genera- tion, Translation, and Comprehension. In ACL. Association for...

    2020

  13. [21]

    Spithourakis, Jianfeng Gao, and William B

    Jiwei Li, Michel Galley, Chris Brockett, Georgios P. Spithourakis, Jianfeng Gao, and William B. Dolan. 2016. A Persona-Based Neural Conversation Model. In ACL (1). The Association for Computer Linguistics

    2016

  14. [22]

    Paul Pu Liang, Terrance Liu, Ziyin Liu, Ruslan Salakhutdinov, and Louis-Philippe Morency. 2020. Think Locally, Act Globally: Federated Learning with Local and Global Representations. CoRR abs/2001.01523 (2020)

    2020 arXiv

  15. [23]

    Shuai Liu, Hyundong Cho, Marjorie Freedman, Xuezhe Ma, and Jonathan May

  16. [24]

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. 2017. Communication-Efficient Learning of Deep Net- works from Decentralized Data. In AISTATS (Proceedings of Machine Learning Research, Vol. 54). PMLR, 1273–1282

    2017

  17. [25]

    Brendan McMahan, Daniel Ramage, Kunal Talwar, and Li Zhang

    H. Brendan McMahan, Daniel Ramage, Kunal Talwar, and Li Zhang. 2018. Learn- ing Differentially Private Recurrent Language Models. In ICLR (Poster). OpenRe- view.net

    2018

  18. [26]

    Yanru Qu, Han Cai, Kan Ren, Weinan Zhang, Yong Yu, Ying Wen, and Jun Wang

  19. [27]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog 1, 8 (2019), 9

    2019

  20. [28]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer.J. Mach. Learn. Res. 21 (2020), 140:1–140:67

    2020

  21. [29]

    Minghuan Tan, Yong Dai, Duyu Tang, Zhangyin Feng, Guoping Huang, Jing Jiang, Jiwei Li, and Shuming Shi. 2022. Exploring and Adapting Chinese GPT to Pinyin Input Method. In ACL (1). Association for Computational Linguistics, 1899–1909

    2022

  22. [30]

    Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. 2020. Progres- sive Layered Extraction (PLE): A Novel Multi-Task Learning (MTL) Model for Personalized Recommendations. In RecSys. ACM, 269–278

    2020

  23. [31]

    Duolingo Team. 2023. Introducing Duolingo Max, a learning experience powered by GPT-4. https://blog.duolingo.com/duolingo-max/

    2023

  24. [32]

    Henry Teicher. 1963. Identifiability of finite mixtures.The annals of Mathematical statistics (1963), 1265–1269

    1963

  25. [33]

    Dijana R Vukovic and Igor M Dujlovic. 2016. Facebook messenger bots and their application for business. In 2016 24th Telecommunications Forum (TELFOR) . IEEE, 1–4

    2016

  26. [34]

    Yuwei Wu, Xuezhe Ma, and Diyi Yang. 2021. Personalized Response Gener- ation via Generative Split Memory Network. In NAACL-HLT. Association for Computational Linguistics, 1956–1970

    2021

  27. [35]

    Mengwei Xu, Feng Qian, Qiaozhu Mei, Kang Huang, and Xuanzhe Liu. 2018. DeepType: On-Device Deep Learning for Input Personalization Service with Minimal Privacy Concern. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol. 2, 4 (2018), 197:1–197:26

    2018

  28. [36]

    Sidney J Yakowitz and John D Spragins. 1968. On the identifiability of finite mixtures. The Annals of Mathematical Statistics 39, 1 (1968), 209–214

    1968

  29. [37]

    Peng Yan and Guodong Long. 2023. Personalization Disentanglement for Feder- ated Learning. In ICME. IEEE, 318–323

    2023

  30. [38]

    Yikai Yan, Chaoyue Niu, Renjie Gu, Fan Wu, Shaojie Tang, Lifeng Hua, Chengfei Lyu, and Guihai Chen. 2022. On-Device Learning for Model Personalization with Large-Scale Cloud-Coordinated Domain Adaption. In KDD. ACM, 2180–2190

    2022

  31. [39]

    Xu Zhang, Yinchuan Li, Wenpeng Li, Kaiyang Guo, and Yunfeng Shao. 2022. Personalized Federated Learning via Variational Bayesian Inference. In ICML (Proceedings of Machine Learning Research, Vol. 162) . PMLR, 26293–26310

    2022

  32. [40]

    Guorui Zhou, Xiaoqiang Zhu, Chengru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep Interest Network for Click-Through Rate Prediction. In KDD. ACM, 1059–1068

    2018

  33. [41]

    Wenchao Zuo, Yuhong Wang, and Yueqing Li. 2019. How to Get to Know Your Customers Better? A Case Analysis of Smartphone Users with Chinese Input Method Based on Baidu Index. In Advances in Usability, User Experience and Assistive Technology: Proceedings of the AHFE 2018 Intern...

    2019

  34. [2016]

    Product-Based Neural Networks for User Response Prediction. In ICDM. IEEE Computer Society, 1149–1154

  35. [2023]

    In ACL (1)

    RECAP: Retrieval-Enhanced Context-Aware Prefix Encoder for Personal- ized Dialogue Response Generation. In ACL (1). Association for Computational Linguistics, 8404–8419

  36. [2024]

    CoRR abs/2403.03952 (2024)

    Bridging Language and Items for Retrieval and Recommendation. CoRR abs/2403.03952 (2024)

    2024 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.