Pith. sign in

REVIEW 4 major objections 5 minor 31 references

RecMind: LLM-Enhanced Graph Neural Networks for Personalized Consumer Recommendations

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a frozen LLM aligned with a LightGCN graph and fused by a per-node gate outperforms strong collaborative, sequential, and LLM-augmented baselines on all eight top-K metrics on Yelp and Amazon-Electronics.

desk verdict A sensible LLM+GNN fusion with a genuinely new gating/alignment combination, but the empirical claims are undercut by a contradictory evaluation protocol and missing artifacts—worth a careful revision, not a reject. read the letter →

arxiv 2509.06286 v1 pith:H6DUHK5W submitted 2025-09-08 cs.LG

Chang Xue , Youwei Lu , Chen Yang , Jinming Xing This is my paper
classification cs.LG
keywords recommendersystemslargelanguagemodelsgraphneuralnetworkscontrastivealignmentgatedfusioncold-startrecommendationLightGCNLoRAadapters
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

RecMind is an architecture for personalized recommendation that combines a graph neural network with a frozen large language model. The paper's central claim is that the two views—text-derived embeddings from titles, attributes, and reviews, and structure-derived embeddings from the user-item interaction graph—should be aligned with a contrastive loss and then fused inside the GNN's message-passing layers by a learned per-node gate. On Yelp and Amazon-Electronics, the paper reports that this design beats strong collaborative, sequential, and LLM-augmented baselines on all eight Recall@K and NDCG@K metrics, with relative gains up to +4.53% on Recall@40 and +4.01% on NDCG@40. If true, it matters because it shows LLM semantic knowledge can help recommendation without using the LLM as a slow monolithic ranker, and that language can be made to dominate precisely where graph data is weakest—cold-start and long-tail items—while graph structure stabilizes ranking elsewhere.

What carries the argument

The load-bearing object is the symmetric cross-modal alignment objective (InfoNCE) combined with an intra-layer gating fusion. The gate γ_v^(l) is a scalar in [0,1] computed by an MLP over the concatenation of the current GNN embedding, the LLM embedding, and a normalized degree feature; the fused state is a convex combination of graph and language embeddings fed into the next LightGCN propagation. The gate is what lets language dominate when degree is low and graph structure dominate when collaborative evidence is strong, and the contrastive loss is what makes the two views comparable before the gate blends them.

What would settle it

Re-run the Yelp and Amazon-Electronics comparisons with full-corpus ranking (all items as candidates) instead of ranking against 100 sampled negatives, and check whether RecMind still beats the strongest baseline on all eight metrics. A second check: on the cold-start subset (degree ≤ 3), remove text metadata and see whether NDCG@20 collapses; if it does not, the claimed language-prior mechanism is not carrying the gain.

Watch

Extended reading notes

Core claim

The paper's discovery is an integration recipe: treat the LLM as a preference prior, not the recommender. A frozen LLM with LoRA adapters maps item titles, attributes, and reviews, plus user review snippets, to embeddings; a LightGCN-style graph encoder maps the interaction graph to embeddings; a symmetric temperature-scaled InfoNCE loss aligns the two views for the same user/item and separates mismatched ones; and a scalar gate, computed from current layer embeddings, the language embedding, and log-degree, blends the views at every message-passing step. The final representation is a learned global mix of the language-aware graph embedding and the language embedding. The paper reports top s

Load-bearing premise

The central comparison assumes that ranking against 100 sampled negatives per user is a faithful proxy for ranking against the full catalog; if that assumption fails, the reported gains may not transfer to true all-item ranking.

Editorial extensions

If this is right

  • Deployment cost stays modest: the LLM is frozen and only small adapters plus the GNN are trained, so item embeddings can be precomputed offline and online ranking is graph propagation plus dot products.
  • Cold-start and long-tail items should benefit most, since the gate gives language more weight when degree is low and graph evidence is weak.
  • Removing either the user-side or item-side contrastive alignment term degrades deep-ranking metrics, implying both alignments are load-bearing for the reported results.
  • The same alignment-plus-gating recipe could attach to other graph encoders, not just LightGCN, making the design portable across backbones.

Reading between the lines

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

  • The evaluation protocol is described as ranking with 100 sampled negatives per user; ranking against the full item corpus could yield smaller margins than the reported +0.47% to +4.53% improvements. A full-corpus ranking experiment would clarify whether the claim transfers to true all-item recommendation.
  • The gate uses a degree feature, so it is implicitly a popularity-aware mechanism; it could be repurposed for explicit popularity-bias control, e.g., by calibrating the gate to down-weight high-degree items when diversity is desired.
  • The two-phase training—warm-up on alignment only, then joint training—suggests that the contrastive loss may be doing much of the work by initializing a shared space. An experiment that keeps alignment but replaces gated fusion with late fusion would isolate where the gains actually come from.
  • Because the language signal comes from reviews and metadata, the approach is sensitive to text quality and token budgets; in domains with noisy or auto-generated text (e.g., voice transcripts), the claimed cold-start benefits could shrink.
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

4 major / 5 minor

Summary. RecMind couples a frozen LLM (with lightweight adapters) that encodes user/item text metadata with a LightGCN-style graph encoder, aligns the two embedding views via a symmetric InfoNCE objective, and fuses them through per-node, per-layer gating during message passing. The paper reports experiments on Yelp and Amazon-Electronics, claiming the best result on all eight Recall@K/NDCG@K metrics, with relative gains up to +4.53% (Recall@40) and +4.01% (NDCG@40) over baselines. Ablations show that removing the LLM prior or either alignment term degrades performance, supporting the proposed design.

Significance. If the empirical claims hold, the architecture is a sensible and potentially practical contribution: it treats the LLM as a frozen preference prior rather than as a monolithic ranker, uses parameter-efficient adapters, and provides a lightweight fusion mechanism that can be deployed with precomputed item embeddings. The contrastive alignment between language and graph views and the ablation results are valuable for the growing line of work combining LLMs with collaborative filtering. The main contribution is therefore an empirical recipe, and the paper would be useful to the recommender-systems community once the evaluation protocol is made precise and reproducible.

major comments (4)
  1. [V.C and IV.E] The evaluation protocol is internally contradictory. Section V.C says metrics are computed using 'all-item ranking with 100 sampled negatives per user', but all-item ranking scores every candidate item, whereas sampled ranking scores only the positive plus 100 sampled negatives. These protocols produce different absolute numbers and can change method rankings, especially if negatives are popularity-aware as stated in Section IV.E. The headline claim that RecMind is best on all eight metrics in Table I therefore rests on an unspecified evaluation setup. The smallest margin, +0.47% on Amazon NDCG@20, is small enough to depend on this choice and on seed variance. Please specify the exact protocol, report full-corpus ranking or justify the sampled protocol with a clear definition, provide variance over multiple runs, and state the negative-sampling distribution used at evaluation time.
  2. [V.A] Section V.A says 'We evaluate on three consumer-focused, text-rich benchmarks' but then names only two: Amazon-Electronics and Yelp. The abstract also reports results on only two datasets. This is a factual mismatch that must be corrected: either a third dataset is missing from the description and results, or the 'three' should be 'two'. This is essential for reproducibility and for assessing the generality of the claimed improvements.
  3. [IV.B, IV.E, V.B] Several implementation details needed to reproduce the method and interpret the gains are absent. The paper does not identify which frozen LLM is used (e.g., model name and size), the LoRA rank/alpha, prompt templates, token budgets, batch size, number of sampled negatives n, queue size, or final hyperparameter values for lambda, tau, beta, and the learning schedule. Without these, the reported results and the ablation comparisons are difficult to verify. Given that the central claim is empirical, these details are load-bearing for reproducibility and for assessing whether the improvements come from the proposed alignment/fusion or from hyperparameter choices.
  4. [V.D and Table I] No statistical significance or variance information is reported. The relative improvements over the strongest baseline are as small as +0.47% (Amazon NDCG@20), which may be within run-to-run noise for recommender benchmarks. Please report standard deviations or confidence intervals across multiple seeds, and state whether the same splits and negative samples were used for all methods. This is particularly important because the paper's main claim is that RecMind 'attains the best results on all eight reported metrics'.
minor comments (5)
  1. [Abstract / V.A] The abstract consistently mentions two datasets, but Section V.A says three. Align the wording throughout.
  2. [V.C] The phrase 'all-item ranking with 100 sampled negatives per user' should be replaced with one of the two actual protocols; the current wording is confusing and makes the reported metrics ambiguous.
  3. [IV.C] Equation (5) defines the user alignment loss; the item-side loss L^I_align is mentioned but not written. A symmetric definition would improve clarity.
  4. [IV.D] In Equation (6), the notation for the concatenated input to the gate MLP could be made clearer, especially the role of the normalized degree feature d~_v and the scalar constant c.
  5. [General] The paper does not provide code or a link to an implementation. For an empirically driven paper, releasing code and exact evaluation scripts would substantially increase confidence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: metrics are computed on chronologically held-out test interactions and self-citations are incidental.

full rationale

RecMind's derivation chain (Eqs. 1-10) defines graph embeddings, LLM embeddings, an InfoNCE alignment loss, a gated fusion inside message passing, and a BPR ranking loss; no reported metric appears as a training input or as a fitted constant. Test numbers in Table I come from a chronological leave-one-out split (Section V.A), so the reported Recall@K/NDCG@K values are out-of-sample and not forced by construction. The ablations in Section V.E use the same held-out protocol and compare full RecMind against variants, which is a legitimate ablation rather than a fitted-input prediction. The phrase 'language-aware by construction' in Section IV.D describes how the fused embedding is built, not a prediction that equals its own input. Self-citations to the authors' prior work ([9], [25], [27]) are non-load-bearing: [9] is a generic sequence-modeling citation, [25] is cited only for dropout as an over-smoothing regularizer, and [27] is cited only for a momentum queue for contrastive negatives; none is used as a uniqueness theorem or as the justification for the central empirical claim. The evaluation-protocol wording in Section V.C ('all-item ranking with 100 sampled negatives per user') is internally contradictory and is a genuine verifiability/correctness concern, but it is not a circularity: the numbers are not defined as their own inputs, and the same ambiguity would apply to all baselines. Under the stated rules, that concern belongs to correctness risk, not to the circularity score. No equation-level reduction, renamed known result, or self-citation chain forces the outcome, so the paper is self-contained with respect to circularity.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The empirical claim rests on domain assumptions about preference graphs and text metadata, on standard ML objectives (LightGCN, InfoNCE, LoRA), and on evaluation choices. No new entities are invented; the learned gate and channel weights are parameters, not hypotheses.

free parameters (5)
  • alignment weight lambda
    Weights the contrastive alignment terms in Eq. (10); tuned on validation, value not reported.
  • InfoNCE temperature tau
    Scales logits in Eq. (5); value not reported.
  • fusion weight alpha
    Learned global scalar in Eq. (9); initialization and value not reported.
  • degree scaling constant c
    Normalizes degree feature in the gate, Eq. (6); hand-chosen, value not reported.
  • embedding dimension d and layers L
    LightGCN capacity; grid searched per dataset, final values not reported.
assumptions (7)
  • domain assumption The bipartite interaction graph encodes user preference (implicit feedback).
    Section III defines the problem; all CF baselines rely on the same premise.
  • domain assumption Item titles, attributes, and reviews contain preference-relevant semantics for users.
    Section IV.B constructs text prompts; this is the motivating assumption of the method.
  • domain assumption Frozen LLM with LoRA adapters produces useful text embeddings in the GNN embedding space after projection (Eq. 4).
    The alignment loss is assumed to make the projected text space compatible with the graph space.
  • domain assumption Chronological leave-one-out with core-5 filtering is a fair proxy for ranking quality.
    Section V.A describes the split; conclusions about cold-start rely on this protocol.
  • domain assumption Ranking against 100 sampled negatives approximates all-item ranking.
    Section V.C states 'all-item ranking with 100 sampled negatives per user', a contradictory premise that the claims depend on.
  • standard math LightGCN layer averaging (Eq. 1-2) yields reliable collaborative embeddings.
    Adopted as backbone from [5]; no modification.
  • standard math InfoNCE with in-batch and momentum negatives aligns views (Eq. 5).
    Standard contrastive objective from [27].

how reviews work

0 comments
Cite this review

Pith. "Pith review of RecMind: LLM-Enhanced Graph Neural Networks for Personalized Consumer Recommendations." pith.science (2026). https://pith.science/paper/H6DUHK5W

@misc{pith2026250906286,
  author       = {Pith},
  title        = {Pith review of: RecMind: LLM-Enhanced Graph Neural Networks for Personalized Consumer Recommendations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H6DUHK5W}},
  note         = {Machine review of arXiv:2509.06286}
}
read the original abstract

Personalization is a core capability across consumer technologies, streaming, shopping, wearables, and voice, yet it remains challenged by sparse interactions, fast content churn, and heterogeneous textual signals. We present RecMind, an LLM-enhanced graph recommender that treats the language model as a preference prior rather than a monolithic ranker. A frozen LLM equipped with lightweight adapters produces text-conditioned user/item embeddings from titles, attributes, and reviews; a LightGCN backbone learns collaborative embeddings from the user-item graph. We align the two views with a symmetric contrastive objective and fuse them via intra-layer gating, allowing language to dominate in cold/long-tail regimes and graph structure to stabilize rankings elsewhere. On Yelp and Amazon-Electronics, RecMind attains the best results on all eight reported metrics, with relative improvements up to +4.53\% (Recall@40) and +4.01\% (NDCG@40) over strong baselines. Ablations confirm both the necessity of cross-view alignment and the advantage of gating over late fusion and LLM-only variants.

Figures

Figures reproduced from arXiv: 2509.06286 by the authors.

Figure 1
Figure 1. RecMind architecture: (1) LLM with adapters encodes text metadata into language embeddings; (2) GNN encodes interaction graph into graph [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 27 canonical work pages

  1. [1]

    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

    work page 2022

  2. [2]

    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

  3. [3]

    Neural graph collaborative filtering,

    X. Wang, X. He, M. Wang, F. Feng, and T.-S. Chua, “Neural graph collaborative filtering,” in Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval, 2019, pp. 165–174

    work page 2019

  4. [4]

    Graph convolutional neural networks for web-scale recommender systems,

    R. Ying, R. He, K. Chen, P. Eksombatchai, W. L. Hamilton, and J. Leskovec, “Graph convolutional neural networks for web-scale recommender systems,” in Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, 2018, pp. 974–983

    work page 2018

  5. [5]

    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

    work page 2020

  6. [6]

    Bpr: Bayesian personalized ranking from implicit feedback,

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

    arXiv 2012

  7. [7]

    Large language models meet collaborative filtering: An efficient all-round llm- based recommender system,

    S. Kim, H. Kang, S. Choi, D. Kim, M. Yang, and C. Park, “Large language models meet collaborative filtering: An efficient all-round llm- based recommender system,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 1395– 1406

    work page 2024

  8. [8]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems, vol. 30, 2017

    2017

Show all 31 references
  1. [9]

    Unifying prediction and explanation in time-series transformers via shapley-based pretraining,

    Q. Cheng, J. Xing, C. Xue, and X. Yang, “Unifying prediction and explanation in time-series transformers via shapley-based pretraining,” in 2025 21st IEEE International Colloquium on Signal Processing & Its Applications (CSPA). IEEE, 2025, pp. 325–330

    2025

  2. [10]

    Selective layer fine-tuning for federated health- care nlp: A cost-efficient approach,

    L. Zhang and Y . Li, “Selective layer fine-tuning for federated health- care nlp: A cost-efficient approach,” in 2025 International Conference on Artificial Intelligence, Computer, Data Sciences and Applications (ACDSA 2025), 2025

    2025

  3. [11]

    S3-rec: Self-supervised learning for sequential recom- mendation with mutual information maximization,

    K. Zhou, H. Wang, W. X. Zhao, Y . Zhu, S. Wang, F. Zhang, Z. Wang, and J.-R. Wen, “S3-rec: Self-supervised learning for sequential recom- mendation with mutual information maximization,” in Proceedings of the 29th ACM international conference on information & knowledge managem...

    2020

  4. [12]

    Contrastive learning for sequential recommendation,

    X. Xie, F. Sun, Z. Liu, S. Wu, J. Gao, J. Zhang, B. Ding, and B. Cui, “Contrastive learning for sequential recommendation,” in 2022 IEEE 38th international conference on data engineering (ICDE). IEEE, 2022, pp. 1259–1273

    2022

  5. [13]

    Simgcl: graph con- trastive learning by finding homophily in heterophily,

    C. Liu, C. Yu, N. Gui, Z. Yu, and S. Deng, “Simgcl: graph con- trastive learning by finding homophily in heterophily,” Knowledge and Information Systems, vol. 66, no. 3, pp. 2089–2114, 2024

  6. [14]

    Xsimgcl: Towards extremely simple graph contrastive learning for recommenda- tion,

    J. Yu, X. Xia, T. Chen, L. Cui, N. Q. V . Hung, and H. Yin, “Xsimgcl: Towards extremely simple graph contrastive learning for recommenda- tion,” IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 2, pp. 913–926, 2023

    2023

  7. [15]

    Joint deep modeling of users and items using reviews for recommendation,

    L. Zheng, V . Noroozi, and P. S. Yu, “Joint deep modeling of users and items using reviews for recommendation,” in Proceedings of the tenth ACM international conference on web search and data mining, 2017, pp. 425–434

    2017

  8. [16]

    Neural attentional rating regression with review-level explanations,

    C. Chen, M. Zhang, Y . Liu, and S. Ma, “Neural attentional rating regression with review-level explanations,” in Proceedings of the 2018 world wide web conference, 2018, pp. 1583–1592

    2018

  9. [17]

    Neural news recommendation with attentive multi-view learning,

    C. Wu, F. Wu, M. An, J. Huang, Y . Huang, and X. Xie, “Neural news recommendation with attentive multi-view learning,” arXiv preprint arXiv:1907.05576, 2019

    1907 arXiv

  10. [18]

    Mmgcn: Multi-modal graph convolution network for personalized recommenda- tion of micro-video,

    Y . Wei, X. Wang, L. Nie, X. He, R. Hong, and T.-S. Chua, “Mmgcn: Multi-modal graph convolution network for personalized recommenda- tion of micro-video,” in Proceedings of the 27th ACM international conference on multimedia, 2019, pp. 1437–1445

    2019

  11. [19]

    Gpt4rec: A generative framework for personalized recommendation and user interests interpretation,

    J. Li, W. Zhang, T. Wang, G. Xiong, A. Lu, and G. Medioni, “Gpt4rec: A generative framework for personalized recommendation and user interests interpretation,” arXiv preprint arXiv:2304.03879, 2023

    2023 arXiv

  12. [20]

    Recom- mendation as instruction following: A large language model empowered recommendation approach,

    J. Zhang, R. Xie, Y . Hou, X. Zhao, L. Lin, and J.-R. Wen, “Recom- mendation as instruction following: A large language model empowered recommendation approach,” ACM Transactions on Information Systems, vol. 43, no. 5, pp. 1–37, 2025

    2025

  13. [21]

    Large language models are zero-shot rankers for recommender systems,

    Y . Hou, J. Zhang, Z. Lin, H. Lu, R. Xie, J. McAuley, and W. X. Zhao, “Large language models are zero-shot rankers for recommender systems,” in European Conference on Information Retrieval. Springer, 2024, pp. 364–381

    2024

  14. [22]

    Recommender systems in the era of large language models (llms),

    Z. Zhao, W. Fan, J. Li, Y . Liu, X. Mei, Y . Wang, Z. Wen, F. Wang, X. Zhao, J. Tang et al., “Recommender systems in the era of large language models (llms),” IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 11, pp. 6889–6907, 2024

    2024

  15. [23]

    Collm: Integrating collaborative embeddings into large language models for rec- ommendation,

    Y . Zhang, F. Feng, J. Zhang, K. Bao, Q. Wang, and X. He, “Collm: Integrating collaborative embeddings into large language models for rec- ommendation,” IEEE Transactions on Knowledge and Data Engineering, 2025

    2025

  16. [24]

    Large language models enhanced collaborative filtering,

    Z. Sun, Z. Si, X. Zang, K. Zheng, Y . Song, X. Zhang, and J. Xu, “Large language models enhanced collaborative filtering,” in Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, 2024, pp. 2178–2188

    2024

  17. [25]

    Multi-view fuzzy graph attention networks for enhanced graph learning,

    J. Xing, D. Luo, Q. Cheng, C. Xue, and R. Xing, “Multi-view fuzzy graph attention networks for enhanced graph learning,” arXiv preprint arXiv:2412.17271, 2024

    2024 arXiv

  18. [26]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen et al., “Lora: Low-rank adaptation of large language models.” ICLR, vol. 1, no. 2, p. 3, 2022

    2022

  19. [27]

    Enhancing link prediction with fuzzy graph attention networks and dynamic negative sampling,

    J. Xing, R. Xing, C. Xue, and D. Luo, “Enhancing link prediction with fuzzy graph attention networks and dynamic negative sampling,” arXiv preprint arXiv:2411.07482, 2024

    2024 arXiv

  20. [28]

    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

  21. [29]

    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,” in Proceedings of the 28th ACM international conference on information and knowledge management, 2019, pp. 1441– 1450

    2019

  22. [30]

    Llmrec: Large language models with graph augmentation for recommendation,

    W. Wei, X. Ren, J. Tang, Q. Wang, L. Su, S. Cheng, J. Wang, D. Yin, and C. Huang, “Llmrec: Large language models with graph augmentation for recommendation,” in Proceedings of the 17th ACM international conference on web search and data mining, 2024, pp. 806–815

    2024

  23. [31]

    Enhancing sequential recommendation via llm-based semantic embedding learning,

    J. Hu, W. Xia, X. Zhang, C. Fu, W. Wu, Z. Huan, A. Li, Z. Tang, and J. Zhou, “Enhancing sequential recommendation via llm-based semantic embedding learning,” in Companion Proceedings of the ACM Web Conference 2024, 2024, pp. 103–111

    2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.