Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Cloud-Device Collaborative Agents for Sequential Recommendation

T0 review · 5 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 A cloud-side LLM and a device-side small model cooperate on sequential recommendation, claiming better accuracy and sub-100ms latency than cloud-only or on-device systems.

desk verdict The planner is a real idea, but the evaluation is apples-to-oranges as reported; the candidate-set mismatch needs to be addressed before the accuracy claims can be taken seriously. read the letter →

arxiv 2509.01551 v1 pith:NI3EFMC2 submitted 2025-09-01 cs.IR

Jing Long , Sirui Huang , Huan Huo , Tong Chen , Hongzhi Yin , Guandong Xu This is my paper
classification cs.IR
keywords cloud-devicecollaborationsequentialrecommendationLLMagentssmalllanguagemodelsprivacy-preservingstrategyplanningcandidateretrieval
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 CDA4Rec, a two-agent architecture for sequential recommendation in which a cloud-side large language model handles planning, semantic modeling, and candidate retrieval while a device-side small language model protects local data, builds a structured user model, and performs final ranking. The central claim is that decomposing the recommendation task into modular sub-tasks and letting the cloud agent generate a personalized execution plan from a privacy-preserving user abstract yields higher accuracy than cloud-only, device-only, and existing cloud-device baselines, while cutting inference latency below 100 ms per request. The design matters because it targets the two bottlenecks that have kept LLM-based recommenders out of real deployment: uploading raw user histories to the cloud is privacy-costly, and running a full model on-device is resource-prohibitive. If the results hold, the framework offers a practical division of labor in which the cloud's reasoning is used without exposing raw behavior, and the device's real-time signals are used without sacrificing model capacity.

What carries the argument

The carrying mechanism is the strategy-planning loop built on a privacy-aware user abstract. The abstract is the paper's central object: a two-part natural-language summary (sanitized query plus behavioral summary) that is meant to be information-preserving for planning while withholding raw identifiers and item-level history. The cloud LLM consumes this abstract and emits a personalized execution plan with numeric coefficients (alpha for query-vs-behavior weighting, beta for semantic-vs-structured fusion), a weighted tag-based retrieval dictionary, and an encoder selection. That plan is what converts a static pipeline into an adaptive one and what makes parallel cloud/device execution possi

What would settle it

Use the same trained framework but feed the cloud planner (a) the raw query and full item-level history and (b) a deliberately degraded abstract. If accuracy under full access clearly exceeds CDA4Rec while degraded abstracts are much worse, the abstract is dropping essential signals, not just identifiers. Separately, run a reconstruction or membership attack on the abstract to see whether specific item IDs or exact queries can be recovered; success would invalidate the privacy premise.

Watch

Extended reading notes

Core claim

CDA4Rec's core claim is that the cloud-device coordination problem in recommendation can be solved by a planning-then-execution split. The device SLM first rewrites the user query into a sanitized form and summarizes the interaction history without item-level detail. The cloud LLM reads only this abstract and outputs a structured plan: weights for query versus behavioral summary in the semantic embedding, a weighted tag-group retrieval dictionary, a choice of structured encoder (SASRec or SURGE) or none, and fusion weights for semantic versus structured embeddings. The cloud then computes the semantic embedding and retrieves candidates while, in parallel, the device builds the structured emb

Load-bearing premise

The load-bearing assumption is that a short natural-language abstract—the sanitized query plus a behavioral summary—can simultaneously preserve enough information for the cloud to plan well and remove enough detail to protect privacy; neither half of that trade-off is proven.

Editorial extensions

If this is right

  • Inference can stay under 100 ms per sample because candidate retrieval and semantic embedding on the cloud run in parallel with structured user modeling on the device.
  • Raw interaction histories and original queries need not leave the device: the cloud sees only the abstract, so the same architecture is compatible with privacy-sensitive deployment.
  • The paper reports the largest relative gains for users with fewer than 10 interactions, suggesting semantic modeling and planning can fill in when sequential patterns are absent.
  • The pipeline is no longer one fixed model: retrieval scope, encoder choice, and fusion weights vary per user and domain based on the cloud-generated plan.

Reading between the lines

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

  • Beyond the paper: the privacy argument is functional, not formal, because no reconstruction or leakage test is reported; a natural extension is an adversarial test that tries to recover specific item IDs or raw query wording from the abstract alone.
  • The same plan-then-execute split could apply to other interactive tasks beyond recommendation, such as on-device assistants or search, where a small local model summarizes context and a large model plans.
  • Because the cloud LLM is frozen and used zero-shot, the framework's ceiling depends on the planner's prompt-following; distilling the planner into a smaller model or training it from feedback could make the approach cheaper and more robust.
  • The tag-group retrieval design assumes items have reliable tag taxonomies; in open catalogs or long-tail domains without clean tags, retrieval would need a different indexing structure.
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

5 major / 5 minor

Summary. The paper proposes CDA4Rec, a cloud-device collaborative framework for sequential recommendation. A device-side SLM generates a privacy-preserving user abstract (sanitized query plus behavioral summary); a cloud-side LLM uses this abstract to plan a personalized strategy, including weights for semantic versus behavioral modeling, candidate retrieval configuration, structured modeling choice, and fusion weights. The cloud performs semantic embedding and retrieval over tag groups, the device performs structured modeling and final ranking over the retrieved candidate set. The paper reports experiments on four datasets (Movie, Music, Sports, POI) comparing accuracy (HR/NDCG) and inference latency against cloud, device, and collaborative baselines, plus ablations, history-length analyses, few-shot evaluation, and a case study. The main claims are consistent accuracy improvements and sub-100 ms latency, with a privacy-preserving design.

Significance. If the experimental evaluation is valid, CDA4Rec would be a useful empirical contribution to cloud-device recommendation, integrating LLM planning with on-device structured modeling and dynamic candidate retrieval. The framework's modular decomposition and strategy-planning idea are interesting, and the ablation studies attempt to separate the contributions of planning, abstraction, retrieval scope, and fusion. However, there are no proofs, formal guarantees, or reproducibility artifacts (code is not mentioned), so the contribution rests entirely on the reported empirical results. The accuracy comparison is called into question by an apparent mismatch between the evaluation protocol and CDA4Rec's candidate-set ranking, and the efficiency claim is weakened by contradictory latency numbers. These issues are load-bearing for the paper's central claims, so the current version does not establish the stated conclusions.

major comments (5)
  1. [IV.A and III.G] The evaluation protocol is inconsistent with CDA4Rec's ranking procedure. Section IV.A states that during testing 'we rank the ground-truth item against all items' (with POI limited to a 5 km radius), but Eq. (6)-(7) in Section III.G compute final scores only for items in the retrieved candidate set C_u, with no mechanism for scoring items outside C_u. Since retrieval sizes are {500, 1000, 2000, 5000} while the full corpora have up to 109k items, ranking within C_u is substantially easier than ranking over the full corpus. The paper neither reports candidate recall nor explains how out-of-candidate ground-truth items are handled. This directly affects all accuracy claims in Table II (e.g., Sports HR@10 0.0304 vs. LSC4Rec 0.0236). Please re-run the comparison with all methods scored over the same candidate restriction, report candidate recall, and include error bars or significance tests.
  2. [Tables III and V] The reported inference times are internally contradictory. Table III gives CDA4Rec POI latency as 42.61 ms and LSC4Rec POI latency as 82.35 ms; Table V gives CDA4Rec POI as 82.35 ms and LSC4Rec POI as 325.05 ms. The same model/domain pair appears twice with different values, and the two tables disagree about which method is faster. Since Section IV.C.2's central efficiency claim rests on these numbers, please clarify the measurement protocol, report hardware/software, and provide a single consistent set of latency results.
  3. [Eq. (8)-(9)] The training loss L_abs = l_CE(π_u, π'_u) is underspecified. The 'plans' π_u and π'_u contain heterogeneous structured elements (continuous weights α and β, tag dictionaries, and discrete model choices), while cross-entropy is defined over probability distributions of token sequences. The paper does not define how plans are tokenized or aligned for the CE computation, nor how the adapter Θ_s ⊕ ϕ_abs is attached. Similarly, L_exp compares SLM and LLM explanation strings with l_CE without specifying decoding/alignment. This makes the abstract-adapter training objective ambiguous and weakens the interpretation of the w/o AA ablation.
  4. [III.B and IV.D.2] The privacy preservation claim is asserted but not quantified. The device sends an abstract generated by an SLM from raw item descriptions and interaction timestamps, and the paper states this is 'privacy-preserving' and that the trade-off is 'worthwhile' (Section IV.D.2). No threat model, formal guarantee, or empirical leakage test is given. Because privacy is a stated motivation and contribution, please specify the adversary model and provide at least a leakage/utility measurement (e.g., attribute-inference success or membership inference) to support the claim.
  5. [II and III.I] There is an internal contradiction about item embeddings. Section II states that embeddings are 'stored as the fixed item representation matrix E_I for downstream training,' but Section III.I and Eq. (10) state that 'CDA4Rec learns item embeddings' through the contrastive loss L_REC, which updates e_i via the scoring function. If embeddings are fixed, Eq. (10) cannot update them; if they are trained, the word 'fixed' in Section II is misleading and the retrieval similarity in Eq. (3) may change during training. Please clarify the training scheme and state which embeddings are used for retrieval.
minor comments (5)
  1. [Table I caption] The caption says 'the statistics of the two datasets,' but Table I lists four datasets.
  2. [IV.G case study] The case study plan says 'Structured User Modeling: {SURGE}' but the explanation immediately says 'SASRec is selected for its ability...'. This is a direct internal inconsistency in the presented example.
  3. [IV.A] The text says 'all baseline methods that require an LLM backbone are implemented using LLaMA-3-8B,' while CDA4Rec uses LLaMA-3.1-8B. The version mismatch should be clarified.
  4. [III.I] The symbol ⊕ used in Eqs. (8)-(9) is not defined; it appears to denote adapter-augmented generation but should be spelled out.
  5. [IV.E] Typo: 'the length of a user user's interaction history' should be 'a user's interaction history.'

Circularity Check

0 steps flagged · score 0.0 of 10

No circular dependency in the derivation chain; the evaluation-scope mismatch is a correctness issue, not circularity.

full rationale

CDA4Rec's derivation chain is self-contained: the semantic embedding (Eq. 2), candidate retrieval (Eq. 3), structured embedding (Eq. 4), fusion (Eq. 5), and final ranking (Eqs. 6–7) are all defined from user inputs and LLM/SLM outputs; no quantity is defined in terms of the reported metrics. The strategy-planning coefficients α and β are per-user outputs of a zero-shot LLM planner, not parameters fitted to test labels. The training losses (Eqs. 8–10) are alignment/contrastive objectives: Eq. 8 trains the abstract adapter to imitate the LLM's full-access plan, Eq. 10 trains item/behavior encoders against interaction labels, and neither bakes the evaluation answer into the inputs. Self-citations ([3], [10], [41], [42]) appear only in related-work context and are not load-bearing. The one notable inconsistency is that Section IV.A says 'During testing, we rank the ground-truth item against all items', while Eq. 7 and Section III.G state final ranking is over the cloud-retrieved candidate set C_u. If CDA4Rec's reported HR/NDCG were computed only over C_u while baselines rank the full corpus, the accuracy comparison would be unfair. However, that is an evaluation-validity/correctness concern, not circularity: the candidate set is retrieved from item embeddings and then re-ranked, and the reported metric is not definitionally equal to an input constant. No circular step is present.

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

No new entities (particles, forces, dimensions) are introduced; the 'user abstract' is a data representation, not an invented entity. The few free parameters are standard training hyperparameters (dim=1024, lr=2e-4, dropout=0.2, etc.) and design choices for tag counts and candidate sizes, neither of which is fitted to test data. The key assumptions are about LLM planning reliability, abstract fidelity and privacy, the implementability of the structured output loss, and the consistency of the item embedding training.

assumptions (4)
  • domain assumption The cloud LLM can produce reliable personalized plans (alpha, beta, tag groups, model choice) from a text abstract without being trained against the ranking objective.
    Section III.C relies on the LLM's zero-shot reasoning about query clarity, behavior sparsity, and domain type; no loss ties planning choices to recommendation accuracy.
  • domain assumption The device-generated abstract is simultaneously privacy-preserving and sufficiently informative for cloud planning.
    Section III.B and Eq. 8 assume the abstract can mimic full-access plans; no privacy audit or leakage test is given.
  • ad hoc to paper A cross-entropy loss can be computed between structured plans (continuous weights, discrete model choices, tag dictionaries) in Eqs. 8-9.
    The paper does not specify the encoding or probability model for l_CE over structured outputs.
  • ad hoc to paper Item embeddings can be used as fixed retrieval representations while also being trained in Eq. 10.
    Section II calls the embedding matrix fixed; Section III.I says item embeddings are learned; the paper does not reconcile this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cloud-Device Collaborative Agents for Sequential Recommendation." pith.science (2026). https://pith.science/paper/NI3EFMC2

@misc{pith2026250901551,
  author       = {Pith},
  title        = {Pith review of: Cloud-Device Collaborative Agents for Sequential Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NI3EFMC2}},
  note         = {Machine review of arXiv:2509.01551}
}
read the original abstract

Recent advances in large language models (LLMs) have enabled agent-based recommendation systems with strong semantic understanding and flexible reasoning capabilities. While LLM-based agents deployed in the cloud offer powerful personalization, they often suffer from privacy concerns, limited access to real-time signals, and scalability bottlenecks. Conversely, on-device agents ensure privacy and responsiveness but lack the computational power for global modeling and large-scale retrieval. To bridge these complementary limitations, we propose CDA4Rec, a novel Cloud-Device collaborative framework for sequential Recommendation, powered by dual agents: a cloud-side LLM and a device-side small language model (SLM). CDA4Rec tackles the core challenge of cloud-device coordination by decomposing the recommendation task into modular sub-tasks including semantic modeling, candidate retrieval, structured user modeling, and final ranking, which are allocated to cloud or device based on computational demands and privacy sensitivity. A strategy planning mechanism leverages the cloud agent's reasoning ability to generate personalized execution plans, enabling context-aware task assignment and partial parallel execution across agents. This design ensures real-time responsiveness, improved efficiency, and fine-grained personalization, even under diverse user states and behavioral sparsity. Extensive experiments across multiple real-world datasets demonstrate that CDA4Rec consistently outperforms competitive baselines in both accuracy and efficiency, validating its effectiveness in heterogeneous and resource-constrained environments.

Figures

Figures reproduced from arXiv: 2509.01551 by the authors.

Figure 1
Figure 1. (a) Example of Cloud Inference. (b) Example of on [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overview of our proposed CDA4Rec. large language model Θl to analyze the item’s textual metadata text(i) (e.g., title, brand, tags): ei = PCA(AVG(Θl(text(i)))). (1) Here, PCA(·) denotes a dimensionality reduction step applied offline, which compresses the original high-dimensional LLM output into the target embedding space R d . The resulting embeddings are stored as the fixed item representation matrix EI for d… view at source ↗
Figure 3
Figure 3. The prompt for abstract generation without referencing specific items. This reasoning process is executed via a prompt-driven interface, where the agent uses a structured instruction prompt shown in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The prompt for strategy planning outputs a weighting coefficient β, where a higher β indicates greater emphasis on the semantic embedding derived from query and behavioral summaries, and 1 − β reflects the weight assigned to the structured embedding constructed from se…
Figure 5
Figure 5. Figure 5: The prompt for explanation generation I. Model Optimization Optimizing CDA4Rec focuses on fine-tuning the lightweight SLM and training item embeddings to align device-side behaviors with cloud-side reasoning. This alignment enables effective abstract generation, interp…
Figure 6
Figure 6. Figure 6: Recommendation performance w.r.t different interac [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: An example of Movie Recommendation. recommendation. These approaches decompose the recom￾mendation process into modular sub-tasks and assign them to specialized LLM agents that collaborate through message passing or shared memory. Our work build this line of research a…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Autonomous Information Seeking: A Roadmap for Agentic Recommender Systems

    cs.IR 2026-07 accept novelty 5.0 of 10

    Agentic recommender systems are organized by agent role (assisted, as-recommender, as-simulator) crossed with autonomy levels L2–L5, yielding a roadmap of architectures, evaluation limits, and open challenges.

Reference graph

Works this paper leans on

44 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    Long-and short-term self-attention network for sequential recommen- dation,

    C. Xu, J. Feng, P. Zhao, F. Zhuang, D. Wang, Y . Liu, and V . S. Sheng, “Long-and short-term self-attention network for sequential recommen- dation,” Neurocomputing, vol. 423, pp. 580–589, 2021

    work page 2021

  2. [2]

    Deeprec: On-device deep learning for privacy-preserving sequential recommendation in mobile commerce,

    J. Han, Y . Ma, Q. Mei, and X. Liu, “Deeprec: On-device deep learning for privacy-preserving sequential recommendation in mobile commerce,” in Proceedings of the Web Conference 2021 , 2021, pp. 900–911

    work page 2021

  3. [3]

    Model-agnostic decentralized collaborative learning for on-device poi recommendation,

    J. Long, T. Chen, Q. V . H. Nguyen, G. Xu, K. Zheng, and H. Yin, “Model-agnostic decentralized collaborative learning for on-device poi recommendation,” in Proceedings of the 46th international ACM SIGIR conference on research and development in information retrieval , 2023, pp. 423–432

    work page 2023

  4. [4]

    Lightgcn: Simplifying and powering graph convolution network for recommenda- tion,

    X. He, K. Deng, X. Wang, Y . Li, Y . Zhang, and M. Wang, “Lightgcn: Simplifying and powering graph convolution network for recommenda- tion,” in Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval , 2020, pp. 639– 648. 12

    work page 2020

  5. [5]

    Graph neural networks for recommender system,

    C. Gao, X. Wang, X. He, and Y . Li, “Graph neural networks for recommender system,” in Proceedings of the fifteenth ACM international conference on web search and data mining , 2022, pp. 1623–1625

    work page 2022

  6. [6]

    Data- efficient fine-tuning for llm-based recommendation,

    X. Lin, W. Wang, Y . Li, S. Yang, F. Feng, Y . Wei, and T.-S. Chua, “Data- efficient fine-tuning for llm-based recommendation,” in Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval , 2024, pp. 365–374

    work page 2024

  7. [7]

    A Federated Framework for LLM-based Recommendation

    J. Zhao, W. Wang, C. Xu, Z. Ren, S.-K. Ng, and T.-S. Chua, “Llm-based federated recommendation,” arXiv preprint arXiv:2402.09959 , 2024

    work page Pith review arXiv 2024

  8. [8]

    Rah! recsys-assistant-human: A human-central recommendation framework with large language models,

    Y . Shu, H. Gu, P. Zhang, H. Zhang, T. Lu, D. Li, and N. Gu, “Rah! recsys-assistant-human: A human-central recommendation framework with large language models,” CoRR, 2023

    work page 2023

Show all 44 references
  1. [9]

    Macrec: A multi- agent collaboration framework for recommendation,

    Z. Wang, Y . Yu, W. Zheng, W. Ma, and M. Zhang, “Macrec: A multi- agent collaboration framework for recommendation,” in Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2024, pp. 2760–2764

    2024

  2. [10]

    On-device next-item recommendation with self-supervised knowledge distillation,

    X. Xia, H. Yin, J. Yu, Q. Wang, G. Xu, and Q. V . H. Nguyen, “On-device next-item recommendation with self-supervised knowledge distillation,” in Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2022, pp. 546– 555

    2022

  3. [11]

    On-device language models: A comprehensive review,

    J. Xu, Z. Li, W. Chen, Q. Wang, X. Gao, Q. Cai, and Z. Ling, “On-device language models: A comprehensive review,” arXiv preprint arXiv:2409.00088, 2024

    2024 arXiv

  4. [12]

    Division-of-thoughts: Harnessing hybrid language model synergy for efficient on-device agents,

    C. Shao, X. Hu, Y . Lin, and F. Xu, “Division-of-thoughts: Harnessing hybrid language model synergy for efficient on-device agents,” in Proceedings of the ACM on Web Conference 2025 , 2025, pp. 1822– 1833

    2025

  5. [13]

    Self-attentive sequential recommenda- tion,

    W.-C. Kang and J. McAuley, “Self-attentive sequential recommenda- tion,” in 2018 IEEE international conference on data mining (ICDM) . IEEE, 2018, pp. 197–206

    2018

  6. [14]

    Sequential recommendation with graph neural networks,

    J. Chang, C. Gao, Y . Zheng, Y . Hui, Y . Niu, Y . Song, D. Jin, and Y . Li, “Sequential recommendation with graph neural networks,” in Proceed- ings of the 44th international ACM SIGIR conference on research and development in information retrieval , 2021, pp. 378–387

    2021

  7. [15]

    Qlora: Efficient finetuning of quantized llms,

    T. Dettmers, A. Pagnoni, A. Holtzman, and L. Zettlemoyer, “Qlora: Efficient finetuning of quantized llms,” Advances in neural information processing systems, vol. 36, pp. 10 088–10 115, 2023

    2023

  8. [16]

    Towards deep conversational recommendations,

    R. Li, S. E. Kahou, H. Schulz, V . Michalski, L. Charlin, and C. Pal, “Towards deep conversational recommendations,” inAdvances in Neural Information Processing Systems 31 (NIPS 2018) , 2018

    2018

  9. [17]

    Music4all: A new music database and its applications,

    I. A. P. Santana, F. Pinhelli, J. Donini, L. Catharin, R. B. Mangolin, V . D. Feltrim, M. A. Domingues et al., “Music4all: A new music database and its applications,” in 2020 International Conference on Systems, Signals and Image Processing (IWSSIP) . IEEE, 2020, pp. 399–404

    2020

  10. [18]

    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

  11. [19]

    Revisiting user mobility and social relationships in lbsns: a hypergraph embedding approach,

    D. Yang, B. Qu, J. Yang, and P. Cudre-Mauroux, “Revisiting user mobility and social relationships in lbsns: a hypergraph embedding approach,” in The world wide web conference , 2019, pp. 2147–2157

    2019

  12. [20]

    Cofirank maximum margin matrix factorization for collaborative ranking,

    M. Weimer, A. Karatzoglou, Q. V . Le, and A. Smola, “Cofirank maximum margin matrix factorization for collaborative ranking,” in Proceedings of the 21st International Conference on Neural Information Processing Systems , ser. NIPS’07. Red Hook, NY , USA: Curran Associates Inc....

    2007

  13. [21]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azhar et al. , “Llama: Open and efficient foundation language models,” arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  14. [22]

    Session- based recommendations with recurrent neural networks,

    B. Hidasi, A. Karatzoglou, L. Baltrunas, and D. Tikk, “Session- based recommendations with recurrent neural networks,” arXiv preprint arXiv:1511.06939, 2015

    2015 arXiv

  15. [23]

    Ra-rec: An efficient id representation alignment framework for llm-based recommendation,

    X. Yu, L. Zhang, X. Zhao, Y . Wang, and Z. Ma, “Ra-rec: An efficient id representation alignment framework for llm-based recommendation,” arXiv preprint arXiv:2402.04527 , 2024

    2024 arXiv

  16. [24]

    Bridging items and language: A transition paradigm for large language model- based recommendation,

    X. Lin, W. Wang, Y . Li, F. Feng, S.-K. Ng, and T.-S. Chua, “Bridging items and language: A transition paradigm for large language model- based recommendation,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 1816– 1826

    2024

  17. [25]

    Next point-of-interest recommendation on resource- constrained mobile devices,

    Q. Wang, H. Yin, T. Chen, Z. Huang, H. Wang, Y . Zhao, and N. Q. Viet Hung, “Next point-of-interest recommendation on resource- constrained mobile devices,” in Proceedings of the Web conference 2020, 2020, pp. 906–916

    2020

  18. [26]

    Prefer: Point-of-interest recommendation with efficiency and privacy- preservation via federated edge learning,

    Y . Guo, F. Liu, Z. Cai, H. Zeng, L. Chen, T. Zhou, and N. Xiao, “Prefer: Point-of-interest recommendation with efficiency and privacy- preservation via federated edge learning,” Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 5, no. 1...

    2021

  19. [27]

    Device-cloud collaborative learning for recommendation,

    J. Yao, F. Wang, K. Jia, B. Han, J. Zhou, and H. Yang, “Device-cloud collaborative learning for recommendation,” in Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining , 2021, pp. 3865–3874

    2021

  20. [28]

    Duet: A tuning-free device- cloud collaborative parameters generation framework for efficient device model generalization,

    Z. Lv, W. Zhang, S. Zhang, K. Kuang, F. Wang, Y . Wang, Z. Chen, T. Shen, H. Yang, B. C. Ooi et al. , “Duet: A tuning-free device- cloud collaborative parameters generation framework for efficient device model generalization,” in Proceedings of the ACM Web Conference 2023, 202...

    2023

  21. [29]

    Collaboration of large language models and small recom- mendation models for device-cloud recommendation,

    Z. Lv, T. Zhan, W. Wang, X. Lin, S. Zhang, W. Zhang, J. Li, K. Kuang, and F. Wu, “Collaboration of large language models and small recom- mendation models for device-cloud recommendation,” arXiv preprint arXiv:2501.05647, 2025

    2025 arXiv

  22. [30]

    A review on matrix factorization techniques in recommender systems,

    R. Mehta and K. Rana, “A review on matrix factorization techniques in recommender systems,” in 2017 2nd International Conference on Communication Systems, Computing and IT Applications (CSCITA) . IEEE, 2017, pp. 269–274

    2017

  23. [31]

    Factorizing personalized markov chains for next-basket recommendation,

    S. Rendle, C. Freudenthaler, and L. Schmidt-Thieme, “Factorizing personalized markov chains for next-basket recommendation,” in Pro- ceedings of the 19th international conference on World wide web , 2010, pp. 811–820

    2010

  24. [32]

    Mv-rnn: A multi- view recurrent neural network for sequential recommendation,

    Q. Cui, S. Wu, Q. Liu, W. Zhong, and L. Wang, “Mv-rnn: A multi- view recurrent neural network for sequential recommendation,” IEEE Transactions on Knowledge and Data Engineering , vol. 32, no. 2, pp. 317–331, 2018

    2018

  25. [33]

    Selfgnn: Self-supervised graph neural networks for sequential recommendation,

    Y . Liu, L. Xia, and C. Huang, “Selfgnn: Self-supervised graph neural networks for sequential recommendation,” in Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2024, pp. 1609–1618

    2024

  26. [34]

    Dynamic graph neural networks for sequential recommendation,

    M. Zhang, S. Wu, X. Yu, Q. Liu, and L. Wang, “Dynamic graph neural networks for sequential recommendation,” IEEE Transactions on Knowledge and Data Engineering , vol. 35, no. 5, pp. 4741–4753, 2022

    2022

  27. [35]

    Time interval aware self-attention for sequential recommendation,

    J. Li, Y . Wang, and J. McAuley, “Time interval aware self-attention for sequential recommendation,” in Proceedings of the 13th international conference on web search and data mining , 2020, pp. 322–330

    2020

  28. [36]

    Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5),

    S. Geng, S. Liu, Z. Fu, Y . Ge, and Y . Zhang, “Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5),” in Proceedings of the 16th ACM conference on recommender systems, 2022, pp. 299–315

    2022

  29. [37]

    Prompt distillation for efficient llm- based recommendation,

    L. Li, Y . Zhang, and L. Chen, “Prompt distillation for efficient llm- based recommendation,” in Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , 2023, pp. 1348–1357

    2023

  30. [38]

    Rah! recsys–assistant–human: A human-centered recommendation framework with llm agents,

    Y . Shu, H. Zhang, H. Gu, P. Zhang, T. Lu, D. Li, and N. Gu, “Rah! recsys–assistant–human: A human-centered recommendation framework with llm agents,” IEEE Transactions on Computational Social Systems , 2024

    2024

  31. [39]

    On generative agents in recommendation,

    A. Zhang, Y . Chen, L. Sheng, X. Wang, and T.-S. Chua, “On generative agents in recommendation,” in Proceedings of the 47th international ACM SIGIR conference on research and development in Information Retrieval, 2024, pp. 1807–1817

    2024

  32. [40]

    Federated rec- ommendation systems,

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

    2020

  33. [41]

    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

  34. [42]

    Decentralized collaborative learning framework for next poi recommendation,

    J. Long, T. Chen, Q. V . H. Nguyen, and H. Yin, “Decentralized collaborative learning framework for next poi recommendation,” ACM Transactions on Information Systems , vol. 41, no. 3, pp. 1–25, 2023

    2023

  35. [43]

    Device-cloud collaborative recommendation via meta controller,

    J. Yao, F. Wang, X. Ding, S. Chen, B. Han, J. Zhou, and H. Yang, “Device-cloud collaborative recommendation via meta controller,” in Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2022, pp. 4353–4362

    2022

  36. [44]

    Intelligent model update strategy for sequential recommendation,

    Z. Lv, W. Zhang, Z. Chen, S. Zhang, and K. Kuang, “Intelligent model update strategy for sequential recommendation,” in Proceedings of the ACM Web Conference 2024 , 2024, pp. 3117–3128

    2024

Pith tools

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