arxiv: 2604.10147 · v1 · submitted 2026-04-11 · 💻 cs.IR · cs.AI

Recognition: unknown

MOSAIC: Multi-Domain Orthogonal Session Adaptive Intent Capture for Prescient Recommendations

Abderaouf Bahi , Mourad Boughaba , Ibtissem Gasmi , Warda Deghmane , Amel Ourici

Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:54 UTC · model grok-4.3

classification 💻 cs.IR cs.AI

keywords multi-domain recommendationsession-based recommendationorthogonal decompositionadversarial trainingdynamic gatingpreference factorizationintent captureuser representation

0 comments

The pith

MOSAIC factorizes multi-domain session preferences into three orthogonal components to improve recommendation accuracy and interpretability.

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

The paper sets out to show that user preferences in sessions crossing multiple domains can be cleanly split into domain-specific signals, signals common across domains, and signals unique to the interaction sequence. This split is achieved with a triple-encoder model that applies domain masking, adversarial training through gradient reversal, alignment, and independence constraints, then uses dynamic gating to combine the parts at each step. A sympathetic reader would care because existing multi-domain systems often blend these signals, which reduces accuracy and makes it hard to see why one recommendation is chosen over another. The experiments on two large real-world benchmarks show consistent gains over prior methods, and ablations confirm each element of the split adds value.

Core claim

The central claim is that user preferences in multi-domain sessions can be explicitly factorized into three orthogonal components—domain-specific, domain-common, and cross-sequence-exclusive—using a triple-encoder architecture. Domain masking objectives and adversarial training via gradient reversal, together with representational alignment and mutual independence constraints, enforce the separation. A dynamic gating mechanism then modulates the contribution of each component at every timestep to produce a unified, temporally adaptive session-level user representation that yields higher recommendation accuracy.

What carries the argument

The triple-encoder architecture, with each encoder tied to one preference type and trained under domain masking, adversarial gradient reversal, alignment, mutual independence constraints, and dynamic gating to produce an adaptive combined representation.

If this is right

The model produces more accurate next-item predictions by weighting domain-specific and cross-domain signals differently at each timestep.
Ablation results establish that domain-specific encoding, domain-common modeling, cross-sequence representation, and dynamic gating each contribute to the observed gains.
The orthogonal decomposition supplies direct views into how preferences unique to one domain interact with those shared across domains.
The resulting session representations transfer more effectively across heterogeneous behavioral domains than blended alternatives.

Where Pith is reading between the lines

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

The same separation could be tested in single-domain settings to check whether the cross-sequence component still adds value when no domain boundaries exist.
Applying the framework to datasets with more than two domains would test whether the three-component split remains sufficient or needs extension.
The orthogonal split might reduce negative transfer when adding a new domain to an existing model by keeping the common component stable.

Load-bearing premise

User preferences can be cleanly separated into three mutually independent orthogonal components by the masking, adversarial, alignment, and independence techniques without losing useful information.

What would settle it

An experiment that measures correlation between the three learned components or shows that removing the independence constraints leaves recommendation accuracy unchanged would falsify the separation claim.

Figures

Figures reproduced from arXiv: 2604.10147 by Abderaouf Bahi, Amel Ourici, Ibtissem Gasmi, Mourad Boughaba, Warda Deghmane.

**Figure 1.** Figure 1: MOSAIC’s General workflow Single-domain sequences. The domain-filtered subsequences retaining only items from domain X or Y : s X u = [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗

**Figure 2.** Figure 2: Baselines classification by category MOSAIC vs. cross-domain baselines (RQ1). Among cross-domain models, C3DSR and C2DSRA2 constitute the strongest baselines due to their recent architectural advances. MOSAIC still outperforms C3DSR by +3.6% / +4.1% / +2.8% in NDCG@10 on Amazon Movie–Book, Movie–Music, and Douban Movie–Book, respectively. These gains are statistically significant (p < 0.05) and are attrib… view at source ↗

read the original abstract

Capturing user intent across heterogeneous behavioral domains stands as a fundamental challenge in session-based recommender systems. Yet, existing multi-domain approaches frequently fail to isolate the distinct contribution of cross-domain interactions from those arising within individual domains, limiting their ability to build rich and transferable user representations. In this work, we propose MOSAIC, a Multi-Domain Orthogonal Session Adaptive Intent Capture framework that explicitly factorizes user preferences into three orthogonal components: domain-specific, domain-common, and cross-sequence-exclusive representations. Our approach employs a triple-encoder architecture, where each encoder is dedicated to one preference type, enforced through domain masking objectives and adversarial training via a gradient reversal layer. Representational alignment and mutual independence constraints are jointly optimized to ensure clean preference separation. Additionally, a dynamic gating mechanism modulates the relative contribution of each component at every timestep, yielding a unified and temporally adaptive session-level user representation. We conduct extensive experiments on two large-scale real-world benchmarks spanning multiple domains and interaction types. The ablation study validates that each component domain-specific encoding, domain-common modeling, cross-sequence representation, and dynamic gating contributes meaningfully to the overall performance. Experimental results demonstrate that MOSAIC consistently outperforms state-of-the-art baselines in recommendation accuracy, while simultaneously providing interpretable insights into the interplay between domain-specific and cross-domain preference signals. These findings highlight the potential of orthogonal preference decomposition as a principled strategy for next-generation multi-domain recommender systems.

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.

Desk Editor's Note private letter to a colleague

MOSAIC adds a triple-encoder with masking, adversarial, and independence constraints to split session preferences into domain-specific, common, and cross-sequence parts, plus dynamic gating, but the gains look incremental and the clean separation is not strongly proven.

read the letter

The paper's core move is to run three separate encoders on the same session data, one for each claimed preference type, then force them apart with domain masking, gradient reversal for adversarial training, alignment losses, and mutual independence terms before feeding a dynamic gate that weights them per timestep. That combination is new enough in the multi-domain session setting and directly targets the problem of mixing cross-domain signals with within-domain ones. The ablation results are presented as showing each piece contributes, which is the right check to run, and the experiments use two large real-world multi-domain datasets with standard next-item prediction metrics. The architecture is described clearly enough that someone could reimplement the encoders and constraints without too much guesswork. The dynamic gate is a practical addition that lets the model adapt within a session rather than using a fixed mix. On the positive side, the work stays focused on a real limitation in current multi-domain recommenders and tries to make the factors interpretable. The soft spots are mostly around the central claim of clean orthogonality. Preferences in real sessions are correlated in messy ways, so the constraints may simply trade off some useful signal for the appearance of separation; the paper would need to show that the orthogonal versions actually preserve more transferable information than a single shared encoder with the same total capacity. The abstract and stress-test description do not include the exact baseline list, statistical tests, or per-domain breakdown, which makes it hard to judge whether the reported outperformance is robust or mostly from extra parameters. Citation pattern looks standard for the subfield but does not appear to engage deeply with recent work on contrastive disentanglement or causal factorization in recsys. This paper is for researchers already working on session-based or cross-domain recommenders who want to experiment with explicit preference factorization. A reader looking for a new architecture to adapt or ablate would find usable ideas here. It deserves a serious referee because the experiments are on real data, the ablation is falsifiable, and the method is reproducible enough to test the orthogonality assumption directly. I would send it for review with requests for more baseline details and a clearer comparison of total model capacity.

Referee Report

2 major / 2 minor

Summary. The paper proposes MOSAIC, a Multi-Domain Orthogonal Session Adaptive Intent Capture framework for session-based recommender systems. It factorizes user preferences into three orthogonal components (domain-specific, domain-common, and cross-sequence-exclusive) via a triple-encoder architecture enforced by domain masking, adversarial training with gradient reversal, representational alignment, mutual independence constraints, and a dynamic gating mechanism for temporally adaptive session representations. Experiments on two large-scale real-world multi-domain benchmarks claim consistent outperformance over state-of-the-art baselines in recommendation accuracy, with ablations showing meaningful contributions from each component and providing interpretable insights into domain-specific vs. cross-domain signals.

Significance. If the empirical results and factorization hold under rigorous validation, the work offers a structured way to disentangle preference signals in heterogeneous multi-domain settings using established techniques (masking, gradient reversal, independence losses, gating). This could improve transferability and interpretability in session-based recsys, addressing a real gap where prior multi-domain methods mix signals. The ablation-based validation of components is a standard strength that supports falsifiability.

major comments (2)

[§4] §4 (Experiments): The claim of consistent outperformance and 'meaningful' ablation contributions lacks reported statistical tests (e.g., paired t-tests or Wilcoxon with p-values), effect sizes, or confidence intervals on the accuracy metrics; without these, it is impossible to assess whether the gains over baselines are reliable or could be due to variance, undermining the central empirical claim.
[§3.2] §3.2 (Method, independence constraints): The mutual independence and orthogonality are enforced via a combination of masking, gradient reversal, alignment, and independence losses, but the manuscript does not demonstrate (via e.g. correlation matrices or mutual information estimates on held-out representations) that the three components remain approximately orthogonal after training; residual dependencies would directly contradict the factorization premise and affect the dynamic gating interpretation.

minor comments (2)

[Abstract / §1] The abstract and §1 omit any equations or loss formulations, forcing readers to infer the exact optimization objective; adding the key loss terms (even summarized) would improve accessibility.
[§4.2] Baseline descriptions in §4.2 are high-level; specifying the exact hyper-parameters, embedding dimensions, and training protocols used for each SOTA method would aid reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major point below and describe the corresponding revisions.

read point-by-point responses

Referee: [§4] §4 (Experiments): The claim of consistent outperformance and 'meaningful' ablation contributions lacks reported statistical tests (e.g., paired t-tests or Wilcoxon with p-values), effect sizes, or confidence intervals on the accuracy metrics; without these, it is impossible to assess whether the gains over baselines are reliable or could be due to variance, undermining the central empirical claim.

Authors: We agree that statistical validation would strengthen the empirical claims. In the revised manuscript we will add paired t-tests (with p-values) and 95% confidence intervals for all reported accuracy metrics across both benchmarks, together with effect sizes for the primary comparisons against baselines. These additions will be placed in the main results tables and discussed in §4. revision: yes
Referee: [§3.2] §3.2 (Method, independence constraints): The mutual independence and orthogonality are enforced via a combination of masking, gradient reversal, alignment, and independence losses, but the manuscript does not demonstrate (via e.g. correlation matrices or mutual information estimates on held-out representations) that the three components remain approximately orthogonal after training; residual dependencies would directly contradict the factorization premise and affect the dynamic gating interpretation.

Authors: We acknowledge the value of post-training verification. While the training objectives explicitly target independence, we will include in the revised §4 an analysis of the learned representations: Pearson correlation matrices and mutual-information estimates computed on held-out sessions for the three component vectors. This will quantify residual dependencies and support the interpretation of the dynamic gating mechanism. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The MOSAIC framework is built from standard components in multi-domain recommendation and disentangled representation learning: a triple-encoder architecture, domain masking, gradient-reversal adversarial training, representational alignment, mutual-independence losses, and dynamic gating. None of these elements reduce by construction to a fitted parameter or self-citation that defines the claimed factorization; the orthogonality is enforced by explicit loss terms whose success is measured externally via ablations and benchmark comparisons. No self-definitional equations, fitted-input predictions, or load-bearing self-citations appear in the derivation chain. The central claim therefore remains falsifiable and independent of its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Insufficient detail in the abstract to enumerate specific free parameters or invented entities; the core domain assumption of clean orthogonal factorization is implicit but not quantified.

axioms (1)

domain assumption User preferences admit a clean factorization into three mutually independent orthogonal components (domain-specific, domain-common, cross-sequence-exclusive).
This factorization is the foundational premise that enables the triple-encoder design and all subsequent constraints.

pith-pipeline@v0.9.0 · 5572 in / 1282 out tokens · 33198 ms · 2026-05-10T15:54:06.585271+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

41 extracted references · 5 canonical work pages · 2 internal anchors

[1]

BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer,

F. Sun, J. Liu, J. Wu, C. Pei, X. Lin, W. Ou, and P. Jiang, “BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer,” inProceedings of the 28th ACM International Con- ference on Information and Knowledge Management (CIKM). ACM, 2019, pp. 1441–1450

2019
[2]

Session-based Recommendations with Recurrent Neural Networks

B. Hidasi, A. Karatzoglou, L. Baltrunas, and D. Tikk, “Session-based recommendations with recurrent neural networks,” inInternational Con- ference on Learning Representations (ICLR), 2016, arXiv:1511.06939

work page internal anchor Pith review arXiv 2016
[3]

Self-attentive sequential recommenda- tion,

W.-C. Kang and J. McAuley, “Self-attentive sequential recommenda- tion,” inProceedings of the 2018 IEEE International Conference on Data Mining (ICDM). IEEE, 2018, pp. 197–206

2018
[4]

Addressing data sparsity and cold- start challenges in recommender systems using advanced deep learning and self-supervised learning techniques,

A. Sharma, M. K. Guptaet al., “Addressing data sparsity and cold- start challenges in recommender systems using advanced deep learning and self-supervised learning techniques,”Journal of Experimental & Theoretical Artificial Intelligence, vol. 37, no. 8, pp. 1421–1451, 2024

2024
[5]

Rethinking cross-domain sequential recommendation under open-world assumptions,

W. XU, Q. WU, and R. W ANG, “Rethinking cross-domain sequential recommendation under open-world assumptions,” OpenReview, 2025, openReview:rafiZzXKaa

2025
[6]

CDRNP: Cross-domain recommendation to cold-start users via neural process,

X. Li, J. Sheng, J. Cao, W. Zhang, Q. Li, and T. Liu, “CDRNP: Cross-domain recommendation to cold-start users via neural process,” in Proceedings of the 17th ACM International Conference on Web Search and Data Mining (WSDM). ACM, 2024, pp. 378–386

2024
[7]

A survey on cross-domain sequential recommendation,

S. Chen, Z. Xu, W. Pan, Q. Yang, and Z. Ming, “A survey on cross-domain sequential recommendation,” inProceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI), 2024, p. 884, arXiv:2401.04971

work page arXiv 2024
[8]

Recommender systems meet large language model agents: A survey,

X. Zhu, Y . Wang, H. Gaoet al., “Recommender systems meet large language model agents: A survey,”Foundations and Trends in Privacy and Security, vol. 7, no. 4, pp. 247–396, 2025

2025
[9]

Study the impact of homomorphic encryption on the accuracy of recommendation systems in e-commerce,

A. Bahi, I. Gasmi, and S. Bentrad, “Study the impact of homomorphic encryption on the accuracy of recommendation systems in e-commerce,” 2023

2023
[10]

Contrastive cross- domain sequential recommendation,

J. Cao, X. Cong, J. Sheng, T. Liu, and B. Wang, “Contrastive cross- domain sequential recommendation,” inProceedings of the 31st ACM International Conference on Information & Knowledge Management (CIKM). ACM, 2022, pp. 138–147

2022
[11]

DREAM: Decoupled representation via extraction attention module and supervised contrastive learning for cross-domain sequential recommender,

X. Ye, Y . Li, and L. Yao, “DREAM: Decoupled representation via extraction attention module and supervised contrastive learning for cross-domain sequential recommender,” inProceedings of the 17th ACM Conference on Recommender Systems (RecSys). ACM, 2023, pp. 479– 490

2023
[12]

Deep learning for smart grid stability in energy transition,

A. Bahi, I. Gasmi, and S. Bentrad, “Deep learning for smart grid stability in energy transition,” inProceedings of the Fourth International Confer- ence on Technological Advances in Electrical Engineering (ICTAEE’23), May 2023

2023
[13]

NaviGNN: Multi-Agent Reinforcement Learning and Graph Neural Network for Sustainable Mobility in Futuristic Smart Cities

A. Bahi and A. Ourici, “An intelligent agent-based simulation of human mobility in extreme urban morphologies,” 2026. [Online]. Available: https://arxiv.org/abs/2507.15143

work page internal anchor Pith review Pith/arXiv arXiv 2026
[14]

Mixed attention network for cross-domain sequen- tial recommendation,

G. Lin, C. Gao, Y . Zheng, J. Chang, Y . Niu, Y . Song, K. Gai, Z. Li, D. Jin, Y . Liet al., “Mixed attention network for cross-domain sequen- tial recommendation,” inProceedings of the 17th ACM International Conference on Web Search and Data Mining (WSDM). ACM, 2024, pp. 405–413

2024
[15]

LLMCDSR: Enhancing cross-domain sequential recom- mendation with large language models,

L. Xuet al., “LLMCDSR: Enhancing cross-domain sequential recom- mendation with large language models,”ACM Transactions on Informa- tion Systems, 2024

2024
[16]

ABXI: Invariant interest adaptation for task-guided cross-domain sequential recommendation,

Q. Bian, M. V . de Carvalho, T. Li, J. Xu, H. Fang, and Y . Ke, “ABXI: Invariant interest adaptation for task-guided cross-domain sequential recommendation,” inProceedings of the ACM on Web Conference 2025 (WWW). ACM, 2025, arXiv:2501.15118

work page arXiv 2025
[17]

Tri- CDR: Triple sequence learning for cross-domain recommendation,

H. Ma, R. Xie, L. Meng, X. Chen, X. Zhang, L. Lin, and J. Zhou, “Tri- CDR: Triple sequence learning for cross-domain recommendation,” in Proceedings of the ACM Web Conference 2023 (WWW). ACM, 2023, pp. 383–394

2023
[18]

Contrastive multi-view interest learning for cross-domain sequential recommendation,

L. Xuet al., “Contrastive multi-view interest learning for cross-domain sequential recommendation,”ACM Transactions on Information Systems, 2024

2024
[19]

An improved cross- domain sequential recommendation model based on intra-domain and inter-domain contrastive learning,

Q. Li, H. Ma, R. Zhang, W. Jin, and Z. Li, “An improved cross- domain sequential recommendation model based on intra-domain and inter-domain contrastive learning,”Complex & Intelligent Systems, 2024

2024
[20]

FairCDR: Transferring fairness and user preferences for cross-domain recommendation,

C. Yang, W. Liu, J. Liu, and P. S. Yu, “FairCDR: Transferring fairness and user preferences for cross-domain recommendation,” inProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining. ACM, 2025

2025
[21]

Hierarchical gating network for cross-domain sequential recommendation,

J. Liet al., “Hierarchical gating network for cross-domain sequential recommendation,”ACM Transactions on Information Systems, 2025

2025
[22]

TCLRec: Temporal-aware and intent contrastive learning for sequential recommendation,

H. Zhanget al., “TCLRec: Temporal-aware and intent contrastive learning for sequential recommendation,”Symmetry, vol. 17, no. 10, p. 1634, 2025

2025
[23]

Cross-domain sequential recom- mendation: An attention and temporal-aware approach,

D.-D. Wang, L. Xu, Y .-F. Yuet al., “Cross-domain sequential recom- mendation: An attention and temporal-aware approach,”International Journal of Machine Learning and Cybernetics, vol. 17, p. 33, 2026

2026
[24]

Graph neural networks in recommender systems: A survey,

S. Wu, F. Sun, W. Zhang, X. Xie, and B. Cui, “Graph neural networks in recommender systems: A survey,”ACM Computing Surveys, vol. 55, no. 5, pp. 1–37, 2022

2022
[25]

Federated graph learning for cross-domain recom- mendation,

Z. Zhanget al., “Federated graph learning for cross-domain recom- mendation,” inAdvances in Neural Information Processing Systems (NeurIPS), 2024, openReview:UBpPOqrBKE

2024
[26]

Sfnn: A secure and diverse recommender system through graph neural network and regularized variational autoencoder,

A. Bahi, I. Gasmi, S. Bentrad, M. W. Azizi, R. Khantouchi, and M. Uzun-Per, “Sfnn: A secure and diverse recommender system through graph neural network and regularized variational autoencoder,” Knowledge-Based Systems, vol. 332, p. 114983, 2025. [Online]. Available: https://doi.org/10.1016/j.knosys.2025.114983

work page doi:10.1016/j.knosys.2025.114983 2025
[27]

Contrastive cross-domain sequential rec- ommendation with attention-aware mechanism,

W. Zhao, B. Li, and X. Mo, “Contrastive cross-domain sequential rec- ommendation with attention-aware mechanism,”Complex & Intelligent Systems, vol. 11, p. 252, 2025

2025
[28]

Learning disentangled representations for recommendation,

J. Ma, C. Zhou, P. Cui, H. Yang, and W. Zhu, “Learning disentangled representations for recommendation,” inAdvances in Neural Information Processing Systems (NeurIPS), 2019, pp. 5711–5722

2019
[29]

Weakly-supervised disentangled representation learning via filter-based adaptive swapping,

Z. Zonget al., “Weakly-supervised disentangled representation learning via filter-based adaptive swapping,” Transactions on Machine Learning Research, 2025, openReview:K69rKKozZU

2025
[30]

Knowledge-guided disentangled representation learning for recommender systems,

Z. Yaoet al., “Knowledge-guided disentangled representation learning for recommender systems,”ACM Transactions on Information Systems, 2021

2021
[31]

Domain-aware self- prompting for cross-domain sequential recommendations with natural language explanations,

T. F. Boka, Z. Niu, T. T. Tewolde, and R. Duma, “Domain-aware self- prompting for cross-domain sequential recommendations with natural language explanations,”Neural Networks, vol. 193, p. 107969, 2026

2026
[32]

Channel-enhanced contrastive cross-domain sequential recommendation,

L. Zhanget al., “Channel-enhanced contrastive cross-domain sequential recommendation,”Data Science and Engineering, 2025

2025
[33]

A multi-view graph contrastive learning framework for cross-domain sequential recommendation,

J. Caoet al., “A multi-view graph contrastive learning framework for cross-domain sequential recommendation,”ACM Transactions on Recommender Systems, 2025

2025
[34]

FedSCOPE: Federated cross-domain sequential recommendation with decoupled contrastive learning and privacy-preserving semantic enhancement,

L. Zhao, Y . Lin, S. Qinet al., “FedSCOPE: Federated cross-domain sequential recommendation with decoupled contrastive learning and privacy-preserving semantic enhancement,”Scientific Reports, vol. 16, p. 7420, 2026

2026
[35]

Federated privacy- preserving for cross-domain sequential recommendation,

S. Chen, Y . Dong, Y . Shang, X. Xu, and X. Lin, “Federated privacy- preserving for cross-domain sequential recommendation,” inArtificial Neural Networks and Machine Learning – ICANN 2025, ser. Lecture Notes in Computer Science, vol. 16070. Springer, 2026, p. 4

2025
[36]

FairCDSR: Fairness-aware cross-domain sequential recommendation via multi-interest transfer and contrastive learning,

Y . Wanget al., “FairCDSR: Fairness-aware cross-domain sequential recommendation via multi-interest transfer and contrastive learning,” IEEE Transactions on Knowledge and Data Engineering, 2025

2025
[37]

Domain-adversarial training of neural networks,

Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Lavi- olette, M. Marchand, and V . Lempitsky, “Domain-adversarial training of neural networks,”Journal of Machine Learning Research, vol. 17, no. 59, pp. 1–35, 2016

2016
[38]

Graph-attention network with adversarial domain alignment for robust cross-domain facial expression recognition,

R. Ghaedi, A. BabaAhmadi, R. Zwiggelaar, X. Fan, and N. Alam, “Graph-attention network with adversarial domain alignment for robust cross-domain facial expression recognition,” inProceedings of Machine Learning Research, vol. 304, 2025, the 17th Asian Conference on Machine Learning (ACML)

2025
[39]

Dual intent view contrastive learning for knowledge aware recommender systems,

J. Guo, Z. Yin, S. Fenget al., “Dual intent view contrastive learning for knowledge aware recommender systems,”Scientific Reports, vol. 15, p. 2133, 2025

2025
[40]

LLM4Rec: A comprehensive survey on the integration of large language models in recommender systems,

Q. Liuet al., “LLM4Rec: A comprehensive survey on the integration of large language models in recommender systems,”Future Internet, vol. 17, no. 6, p. 252, 2025

2025
[41]

A survey on LLM-powered agents for recommender systems,

Q. Peng, H. Liu, H. Huang, J. Yang, Q. Yang, and M. Shao, “A survey on LLM-powered agents for recommender systems,” inFindings of the Association for Computational Linguistics: EMNLP 2025. Association for Computational Linguistics, 2025, p. 620

2025