Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Optimizing Recall or Relevance? A Multi-Task Multi-Head Approach for Item-to-Item Retrieval in Recommendation

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read MTMH claims that item-to-item retrieval can jointly optimize co-engagement recall and semantic relevance through a multi-task loss and a two-head serving merge, with a tunable quota that adjusts the trade-off without retraining.

desk verdict Solid industrial I2I paper with a useful serving-time alpha knob, but the semantic-relevance claim rests on a metric aligned with its own teacher and needs independent validation. read the letter →

arxiv 2506.06239 v1 pith:KHMDEYHI submitted 2025-06-06 cs.IR

Jiang Zhang , Sumit Kumar , Wei Chang , Yubo Wang , Feng Zhang , Weize Mao , Hanchao Yu , Aashu Singh
show 2 more authors
Min Li Qifan Wang
This is my paper
classification cs.IR
keywords Item-to-itemretrievalSemanticrelevanceRecall-relevancetrade-offMulti-headmulti-tasklearningKnowledgedistillationCo-engagementContentencoderRecommendationsystem
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

The paper argues that item-to-item (I2I) retrieval models trained only on co-engagement data overfit short-term engagement patterns and sacrifice semantic relevance, while models built purely on content embeddings have high semantic relevance but almost no recall. It proposes MTMH, a multi-task, multi-head architecture that keeps both objectives: an InfoNCE co-engagement loss and a KL-divergence knowledge-distillation loss from a pre-trained multimodal content encoder, with separate engagement and relevance heads whose candidates are merged at serving time by a tunable quota. On proprietary data from a commercial platform, MTMH reports recall@500 of 17.02% and L2 topic relevance of 28.02%, improving recall by up to 14.4% and semantic relevance by up to 56.5% over baselines. A seven-day online A/B test reports gains in daily active users, time spent, distinct item views, fresh content share, novel interest discovery, and user interest recall. If correct, the method gives recommendation systems a serving-time knob to tune the recall-relevance trade-off without retraining.

What carries the argument

The load-bearing object is the two-head multi-task architecture with a quota-based merge. The engagement head optimizes only the InfoNCE co-engagement loss $\mathcal{L}_e$ (Eq. 1). The relevance head optimizes $\mathcal{L}_{mt} = \mathcal{L}_e + w_r \mathcal{L}_r$, where $\mathcal{L}_r$ is the KL divergence $\sum_i \sum_j D_{KL}(Q_{i,j} \| P_{i,j})$ between the teacher's soft similarity distribution $Q$ (from content embeddings $F$) and the student's distribution $P$ (from learned embeddings $E$), implementing knowledge distillation from a pre-trained multimodal content encoder trained by contrastive learning on hashtags, search queries, and LLM-generated tags. During serving, the multi-ANN module retrieves candidates from each head in parallel, the preranker reranks each head's candidates separately, and the merging module keeps the top $\alpha\%$ of engagement-head candidates and $(100-\alpha)\%$ of relevance-head candidates, so $\alpha$ acts as a recall-relevance knob that can be turned without retraining.

What would settle it

Train the same two-head model with the relevance-loss weight set to zero ($w_r = 0$). If its recall@500 and L2 topic relevance match the full MTMH numbers, then the multi-task relevance loss is doing none of the reported work and the gains come from the merge alone; if a sweep over $w_r$ never lifts recall@500 above the pure engagement head, then the method trades off rather than jointly optimizes. A second check is to run the same architecture on a public recommendation dataset with an independently trained content encoder: if the $\alpha$ sweep does not trace a Pareto frontier that dominates the co-engagement baseline, the results are specific to the proprietary teacher and platform rather than to the method.

Watch

Extended reading notes

Core claim

The central claim is that the recall-relevance trade-off in I2I retrieval is not a fixed constraint: a multi-task objective can pull learned item embeddings toward both co-engagement and content semantics, and a multi-head serving strategy can then harvest both kinds of candidates. The co-engagement head is trained only on the InfoNCE loss, maximizing the probability that a trigger item identifies positively co-engaged candidates among negatives. The relevance head is trained on the weighted sum of that loss and a KL-divergence term that aligns the model's similarity distribution over candidate items with the similarity distribution produced by a pre-trained content encoder (the teacher). At serving time, MTMH retrieves $K_{ann}$ candidates per head via ANN, reranks each head's candidates separately, and merges the top $\alpha\%$ from the engagement head with the top $(100-\alpha)\%$ from the relevance head, removing duplicates. In the paper's experiments this yields recall@500 of 17.02% versus 16.33% for the best single-objective baseline, while L2 topic relevance rises to 28.02% versus 26.25%, and online A/B metrics move in the same direction.

Load-bearing premise

The whole semantic-relevance gain rests on the assumption that the pre-trained content encoder's notion of item similarity, learned from hashtags, search queries, and LLM-generated tags, is a faithful teacher for what users actually find relevant, and that the human-labeled topic match rate used for evaluation is the right external target.

Editorial extensions

If this is right

  • Deploying MTMH with $\alpha$ adjustable at serving time lets operators tune the recall-relevance trade-off per application, content type, or user segment without retraining the model.
  • Because the relevance head distills content-encoder knowledge into lightweight retrieval embeddings during training, the serving model can carry semantic relevance without running a large multimodal encoder online.
  • MTMH accelerates embedding convergence for fresh content, so new items can enter retrieval candidate sets with less engagement history.
  • Live A/B results indicate the approach improves not only consumption metrics (DAU, time spent, distinct item views) but also user-experience metrics such as novel interest discovery and user interest recall.

Reading between the lines

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

  • Beyond the paper, the same multi-task plus dual-head pattern could be applied to user-to-item retrieval or search candidate generation, where the same co-engagement-versus-semantics tension exists.
  • The $\alpha$ knob effectively defines a Pareto frontier over recall and relevance; a natural next step the authors leave implicit is to learn $\alpha$ per user or per context rather than setting it globally.
  • The reported long-term gains come from a seven-day online experiment; a testable extension would measure whether the novelty and diversity gains persist beyond the first month, since the co-engagement head may pull the system back toward popular items over time.
  • If the content encoder's teacher signal is biased toward topics in its training labels, the relevance head could inherit that bias; a platform-specific calibration of the teacher against downstream engagement would test this.
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 / 4 minor

Summary. The paper proposes MTMH, a multi-task multi-head architecture for item-to-item (I2I) retrieval. One head (engagement) is trained with an InfoNCE co-engagement loss; the other (relevance) is trained with a multi-task loss that adds KL distillation from a fixed pretrained multimodal content encoder. During serving, the top candidates from the two heads are merged with a tunable quota alpha. Offline evaluation on proprietary data reports recall@500 of 17.02% (up to 14.4% relative gain) and L2 topic relevance of 28.02% (up to 56.5% relative gain). A 7-day A/B test reports small gains in DAU, time spent, distinct item views, fresh content, novel interest discovery, and user interest recall.

Significance. The central idea of separating an engagement head from a relevance head and merging at serving time is practically useful: alpha provides a training-free trade-off knob. The paper demonstrates internal consistency: the losses are clearly defined in Eqs. (1)-(3), the ablation in Sec. 4.3 isolates the contributions of the multi-task loss and the multi-head design, and the offline trade-off curves in Sec. 4.4 are falsifiable. However, the semantic-relevance claim rests on a topic-match metric that is conceptually aligned with the teacher signal used in the distillation loss, and all results are single-run point estimates or a short A/B test without significance values. If the claims hold, this is a solid industrial contribution; at present the external validity of the relevance claim is not fully established.

major comments (3)
  1. [Sec. 2 and Sec. 3.4 (Eq. 2)] The semantic relevance metric used throughout the paper is the human-labeled topic match rate (Sec. 2), while the pre-trained content encoder used as the distillation teacher in Eq. (2) is trained via contrastive learning on user hashtags, search queries, and LLM-generated tags (Sec. 3.4). Because the teacher and the evaluation metric are driven by the same content-semantics signal, the observed L2 topic relevance improvements in Table 1 may largely reflect alignment with the teacher's notion of similarity rather than user-perceived relevance. The paper does not provide an independent validation of semantic relevance (e.g., graded human relevance judgments on the retrieved candidates, or an external benchmark not based on topic labels). The online A/B metrics in Sec. 4.6 are not defined in enough detail and are reported without significance values, so they do not currently resolve this concern. Please add an independent relevance evaluation or explicitly justify why topic match is the right ground truth and report inter-annotator agreement.
  2. [Sec. 4.2 (Table 1) and Sec. 4.6 (Table 5)] All offline results are point estimates on a single evaluation set, with no confidence intervals, no multiple seeds, and no repeated experiments; the online A/B test is a single 7-day experiment with metric changes of 0.05% to 0.33% and no significance levels, confidence intervals, or definitions of variance. Since the central claims are quantitative (up to 14.4% recall and 56.5% relevance improvements), the paper should report uncertainty quantifications for both offline and online experiments and clarify whether each reported change is statistically significant. This is essential for assessing the magnitude claims in the abstract and the conclusions.
  3. [Sec. 4.1 (Table 1 note)] The baselines MoL, HLLM, and HSTU* are described as modified versions of prior work adapted for item-to-item retrieval, but the precise adaptation (e.g., which content features are used, how HSTU item embeddings are extracted, and whether the original training objectives are preserved) is not specified. Because these baselines are the basis for the claimed improvements over prior SOTA, the paper should describe the adaptations concretely or make the baseline code available, so that the comparison can be reproduced.
minor comments (4)
  1. [Sec. 4.6, Table 5] The header contains a typo: "Perentage of fresh content" should read "Percentage of fresh content."
  2. [Abstract] The abstract reports "recall by up to 14.4%" and "semantic relevance by up to 56.6%" but these relative gains are against different baselines in Table 1; the abstract should state the baseline for each claim to avoid overstatement.
  3. [Eq. (2)] The definitions of q^+_ij and q^-_ik share the same denominator and could be simplified; also p^-_ik is defined only after Eq. (2), which may confuse readers. Consider reordering the notation.
  4. [Sec. 4.5, Figure 7] The y-axis labeled "embedding delta" would benefit from an explicit definition of the L2 distance and the time range over which convergence is measured.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the relevance teacher and the evaluation metric are distinct, and the recall metric is independent future-engagement data.

full rationale

The derivation chain is self-contained with respect to circularity concerns. MTMH's relevance head is trained by KL distillation (Eq. 2) toward soft labels produced by a pre-trained content encoder (Sec. 3.4) that was itself trained by contrastive learning on hashtags, search queries, and LLM-generated tags. The offline semantic-relevance metric is the average L1/L2 topic-category match rate from human labels (Secs. 2 and 4.1). These are not the same object: the teacher's training signal is not the human topic label used in evaluation, and no equation in the paper defines topic match as the content-encoder similarity. Thus the reported relevance gains are an empirical transfer result, not a tautology. Recall@K is measured against future user interaction history, which is independent of the co-engagement training pairs. The only self-citation of note (Ref. [59], which shares a co-author) is used to weight semantically relevant positive pairs via SIM; it is not load-bearing for the central recall/relevance claim, which is separately supported by offline held-out metrics and online A/B results. Lack of confidence intervals or significance tests for the online metrics is a reporting/validity concern, not a circularity. Therefore no circular step can be exhibited under the required standard.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard ML assumptions about co-engagement data, negative sampling, ANN recall, and the quality of the content-encoder teacher. The only fitted hyperparameters that directly control the reported trade-off are w_r and alpha, both selected on the evaluation set. No new physical or conceptual entities are postulated; the content encoder and two-head architecture are models, not entities in the sense of new particles, forces, or conserved quantities.

free parameters (5)
  • w_r = 0.5 (default, tuned on eval)
    Weight of relevance loss in Eq. (3); chosen from Table 4 to balance recall and relevance, with values 0, 0.25, 0.5, 1, 5 evaluated.
  • alpha = 50 (default, tuned on eval)
    Percentage of candidates merged from engagement head in Section 3.5; chosen from Figure 6b as the best trade-off point.
  • K_ann = O(1000)
    Number of ANN candidates per head in serving; chosen by hand, not justified in the paper.
  • K = O(10)
    Number of preranked candidates preserved per head; chosen by hand, not justified.
  • SIM weighting = unspecified
    Positive pairs are weighted more heavily when SIM identifies semantic relevance (Section 3.1); the weighting function is not specified and is treated as a free design choice.
assumptions (6)
  • domain assumption Future user engagement in period T2 is valid ground truth for I2I recall.
    Section 2 evaluation setup: recall@500 measures whether items engaged in the future are retrieved from past triggers.
  • domain assumption Topic match rate between trigger and candidate, from human labels, is a valid measure of semantic relevance.
    Section 2: semantic relevance is defined as the average L1/L2 topic match rate; all relevance claims depend on this proxy.
  • domain assumption The pre-trained content encoder produces a meaningful teacher distribution for semantic similarity.
    Section 3.4: the content encoder is trained with contrastive learning on hashtags, search queries, and LLM-generated tags, and is used to generate soft labels in Eq. (2).
  • domain assumption Randomly sampled negative pairs represent non-relevant items sufficiently well.
    Section 3.1: negative pairs are random non-engaged items; the InfoNCE loss (Eq. 1) depends entirely on this negative distribution.
  • domain assumption ANN search over K-means clusters returns approximately the true nearest neighbors in each embedding space.
    Section 3.5: serving uses C nearest clusters and top K_ann candidates; the final candidates are a subset of true ANN results.
  • domain assumption The preranker U2I model ranks candidates within each head without introducing bias that changes the relative head comparison.
    Section 3.5: the preranker is a user-to-item model shared across heads; the paper assumes it preserves enough candidates from each head for the quota merge to work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimizing Recall or Relevance? A Multi-Task Multi-Head Approach for Item-to-Item Retrieval in Recommendation." pith.science (2026). https://pith.science/paper/KHMDEYHI

@misc{pith2026250606239,
  author       = {Pith},
  title        = {Pith review of: Optimizing Recall or Relevance? A Multi-Task Multi-Head Approach for Item-to-Item Retrieval in Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KHMDEYHI}},
  note         = {Machine review of arXiv:2506.06239}
}
read the original abstract

The task of item-to-item (I2I) retrieval is to identify a set of relevant and highly engaging items based on a given trigger item. It is a crucial component in modern recommendation systems, where users' previously engaged items serve as trigger items to retrieve relevant content for future engagement. However, existing I2I retrieval models in industry are primarily built on co-engagement data and optimized using the recall measure, which overly emphasizes co-engagement patterns while failing to capture semantic relevance. This often leads to overfitting short-term co-engagement trends at the expense of long-term benefits such as discovering novel interests and promoting content diversity. To address this challenge, we propose MTMH, a Multi-Task and Multi-Head I2I retrieval model that achieves both high recall and semantic relevance. Our model consists of two key components: 1) a multi-task learning loss for formally optimizing the trade-off between recall and semantic relevance, and 2) a multi-head I2I retrieval architecture for retrieving both highly co-engaged and semantically relevant items. We evaluate MTMH using proprietary data from a commercial platform serving billions of users and demonstrate that it can improve recall by up to 14.4% and semantic relevance by up to 56.6% compared with prior state-of-the-art models. We also conduct live experiments to verify that MTMH can enhance both short-term consumption metrics and long-term user-experience-related metrics. Our work provides a principled approach for jointly optimizing I2I recall and semantic relevance, which has significant implications for improving the overall performance of recommendation systems.

Figures

Figures reproduced from arXiv: 2506.06239 by the authors.

Figure 1
Figure 1. Recall vs. Relevance for I2I retrieval models. The [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Recall vs relevance for I2I retrieval models. The [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed MTMH approach. 1) The multi-task learning consists of a co-engagement loss and a relevance [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: MTMH serving pipeline contains three modules: [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: Recall and semantic relevance trade-off evaluation [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: MTMH embedding convergence of fresh content. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

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. Full citation record

  1. CAPTS: Channel-Aware, Preference-Aligned Trigger Selection for Multi-Channel Item-to-Item Retrieval

    cs.IR 2026-02 conditional novelty 6.0 of 10

    CAPTS selects retrieval triggers by predicted downstream engagement per I2I channel and routes them jointly, reporting offline Recall@K gains and a small online time-spent lift.

Reference graph

Works this paper leans on

72 extracted references · 36 canonical work pages · cited by 1 Pith paper

  1. [1]

    Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. Llm2vec: Large language models are secretly powerful text encoders.arXiv preprint arXiv:2404.05961(2024)

    arXiv 2024

  2. [2]

    Artun Boz, Wouter Zorgdrager, Zoe Kotti, Jesse Harte, Panagiotis Louridas, Diet- mar Jannach, and Marios Fragkoulis. 2024. Improving Sequential Recommenda- tions with LLMs.ArXivabs/2402.01339 (2024). https://api.semanticscholar.org/ CorpusID:267406555

    arXiv 2024

  3. [3]

    Junyi Chen, Lu Chi, Bingyue Peng, and Zehuan Yuan. 2024. Hllm: Enhancing sequential recommendations via hierarchical large language models for item and user modeling.arXiv preprint arXiv:2409.12740(2024)

    arXiv 2024

  4. [4]

    Jiawei Chen, Hande Dong, Xiang Wang, Fuli Feng, Meng Wang, and Xiangnan He. 2023. Bias and debias in recommender system: A survey and future directions. ACM Transactions on Information Systems41, 3 (2023), 1–39

    2023

  5. [5]

    Lei Chen, Chen Gao, Xiaoyi Du, Hengliang Luo, Depeng Jin, Yong Li, and Meng Wang. 2024. Enhancing ID-based Recommendation with Large Language Mod- els.ArXivabs/2411.02041 (2024). https://api.semanticscholar.org/CorpusID: 273812191

    work page Pith review arXiv 2024

  6. [6]

    Zhang, and Shenghe Li

    Zhixuan Chu, Hongyan Hao, Ouyang Xin, Simeng Wang, Yan Wang, Yue Shen, Jinjie Gu, Qing Cui, Longfei Li, Siqiao Xue, James Y. Zhang, and Shenghe Li. 2023. Leveraging Large Language Models for Pre-trained Recommender Systems.ArXiv abs/2308.10837 (2023). https://api.semanticscholar.org/CorpusID:261049176

    arXiv 2023

  7. [7]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep neural networks for youtube recommendations. InProceedings of the 10th ACM Conference on Recommender Systems. 191–198

    work page 2016

  8. [8]

    Sunhao Dai, Ninglu Shao, Haiyuan Zhao, Weijie Yu, Zihua Si, Chen Xu, Zhongx- iang Sun, Xiao Zhang, and Jun Xu. 2023. Uncovering chatgpt’s capabilities in KDD ’25, August 3–7, 2025, Toronto, ON, Canada. Jiang Zhang et al. recommender systems. InProceedings of the 17th ACM Conference on Recom- mender Systems. 1126–1132

    work page 2023

Show all 72 references
  1. [9]

    Mukund Deshpande and George Karypis. 2004. Item-based top-n recommenda- tion algorithms.ACM Transactions on Information Systems (TOIS)22, 1 (2004), 143–177

    2004

  2. [10]

    Wenqi Fan, Zihuai Zhao, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Jiliang Tang, and Qing Li. 2023. Recommender Systems in the Era of Large Language Models (LLMs).IEEE Transactions on Knowledge and Data Engineering36 (2023), 6889–6907. https://api.semanticscholar.org/Cor...

    2023

  3. [11]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5).Proceedings of the 16th ACM Conference on Rec- ommender Systems(2022). https://api.semanti...

    2022

  4. [12]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. InProceedings of the 26th international conference on world wide web. 173–182

    2017

  5. [13]

    Balázs Hidasi and Alexandros Karatzoglou. 2017. Recurrent Neural Networks with Top-k Gains for Session-based Recommendations.Proceedings of the 27th ACM International Conference on Information and Knowledge Management(2017). https://api.semanticscholar.org/CorpusID:1159769

    2017

  6. [14]

    Balázs Hidasi and Domonkos Tikk. 2013. Context-aware item-to-item recommen- dation within the factorization framework. InProceedings of the 3rd Workshop on Context-awareness in Retrieval and Recommendation. 19–25

    2013

  7. [15]

    Geoffrey Hinton. 2015. Distilling the Knowledge in a Neural Network.arXiv preprint arXiv:1503.02531(2015)

    2015 arXiv

  8. [16]

    Yifan Hu, Yehuda Koren, and Chris Volinsky. 2008. Collaborative filtering for implicit feedback datasets. In2008 Eighth IEEE International Conference on Data Mining. 263–272

    2008

  9. [17]

    Po-Sen Huang, Xiaodong He, Jianfeng Gao, Li Deng, Alex Acero, and Larry Heck. 2013. Learning deep structured semantic models for web search using clickthrough data. InProceedings of the 22nd ACM International Conference on Information & Knowledge Management. 2333–2338

    2013

  10. [18]

    Amir H Jadidinejad, Craig Macdonald, and Iadh Ounis. 2020. Using exploration to alleviate closed loop effects in recommender systems. InProceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. 2025–2028

    2020

  11. [19]

    Ziyan Jiang, Rui Meng, Xinyi Yang, Semih Yavuz, Yingbo Zhou, and Wenhu Chen. 2024. Vlm2vec: Training vision-language models for massive multimodal embedding tasks.arXiv preprint arXiv:2410.05160(2024)

    2024 arXiv

  12. [20]

    Mohammad Khoshneshin and W Nick Street. 2010. Collaborative filtering via eu- clidean embedding. InProceedings of the fourth ACM conference on Recommender systems. 87–94

    2010

  13. [21]

    Mohammad Khoshneshin and William Nick Street. 2010. Collaborative filtering via euclidean embedding. InACM Conference on Recommender Systems. https: //api.semanticscholar.org/CorpusID:7176266

    2010

  14. [22]

    Yehuda Koren. 2008. Factorization meets the neighborhood: a multifaceted collaborative filtering model. InKnowledge Discovery and Data Mining. https: //api.semanticscholar.org/CorpusID:207168823

    2008

  15. [23]

    Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization tech- niques for recommender systems.Computer42, 8 (2009), 30–37

    2009

  16. [24]

    Chao Li, Zhiyuan Liu, Mengmeng Wu, Yuchi Xu, Huan Zhao, Pipei Huang, Guoliang Kang, Qiwei Chen, Wei Li, and Dik Lun Lee. 2019. Multi-interest network with dynamic routing for recommendation at Tmall. InProceedings of the 28th ACM International Conference on Information and Kno...

    2019

  17. [25]

    Lei Li, Yongfeng Zhang, Dugang Liu, and L. Chen. 2023. Large Language Models for Generative Recommendation: A Survey and Visionary Discussions. InInternational Conference on Language Resources and Evaluation. https: //api.semanticscholar.org/CorpusID:261531422

    2023

  18. [26]

    Xiangyang Li, Bo Chen, Huifeng Guo, Jingjie Li, Chenxu Zhu, Xiang Long, Sujian Li, Yichao Wang, Wei Guo, Longxia Mao, Jinxing Liu, Zhenhua Dong, and Ruiming Tang. 2022. IntTower: The Next Generation of Two-Tower Model for Pre-Ranking System.Proceedings of the 31st ACM Internat...

    2022

  19. [27]

    Xinyi Li, Yongfeng Zhang, and Edward C Malthouse. 2023. Pbnr: Prompt-based news recommender system.arXiv preprint arXiv:2304.07862(2023)

    2023 arXiv

  20. [28]

    Dawen Liang, Laurent Charlin, James McInerney, and David M Blei. 2016. Mod- eling user exposure in recommendation. InProceedings of the 25th International Conference on World Wide Web. 951–961

    2016

  21. [29]

    Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Xiangyang Li, Chenxu Zhu, Huifeng Guo, Yong Yu, Ruiming Tang, and Weinan Zhang. 2023. How Can Recommender Systems Benefit from Large Language Models: A Survey.ArXiv abs/2306.05817 (2023). https://api.semanticscholar.org/...

    2023 arXiv

  22. [30]

    Linden, B

    G. Linden, B. Smith, and J. York. 2003. Amazon.com recommendations: item-to- item collaborative filtering.IEEE Internet Computing7, 1 (2003), 76–80. doi:10. 1109/MIC.2003.1167344

    2003 arXiv

  23. [31]

    Qidong Liu, Xiangyu Zhao, Yuhao Wang, Yejing Wang, Zijian Zhang, Yuqi Sun, Xiang Li, Maolin Wang, Pengyue Jia, Chong Chen, Wei Huang, and Feng Tian

  24. [32]

    Zheng Liu, Jianxun Lian, Junhan Yang, Defu Lian, and Xing Xie. 2020. Octopus: Comprehensive and elastic user representation for the generation of recommen- dation candidates. InProceedings of the 43rd International ACM SIGIR Conference on Research and Development in Informatio...

    2020

  25. [33]

    Pasquale Lops, Marco Degemmis, and Giovanni Semeraro. 2011. Content-based Recommender Systems: State of the Art and Trends. InRecommender Systems Handbook. https://api.semanticscholar.org/CorpusID:6102334

    2011

  26. [34]

    Junmei Lv, Bin Song, Jie Guo, Xiaojiang Du, and Mohsen Guizani. 2019. Interest- related item similarity model based on multimodal data for top-N recommenda- tion.IEEE access7 (2019), 12809–12821

    2019

  27. [35]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748(2018)

    2018 arXiv

  28. [36]

    Qi Pi, Guorui Zhou, Yujing Zhang, Zhe Wang, Lejian Ren, Ying Fan, Xiaoqiang Zhu, and Kun Gai. 2020. Search-based user interest modeling with lifelong sequential behavior data for click-through rate prediction. InProceedings of the 29th ACM International Conference on Informati...

    2020

  29. [37]

    Steffen Rendle. 2010. Factorization machines. In2010 IEEE International Confer- ence on Data Mining. 995–1000

    2010

  30. [38]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme

  31. [39]

    Badrul Sarwar, George Karypis, Joseph Konstan, and John Riedl. 2001. Item-based collaborative filtering recommendation algorithms. InProceedings of the 10th International Conference on World Wide Web. 285–295

    2001

  32. [40]

    Konstan, and John Riedl

    Badrul Munir Sarwar, George Karypis, Joseph A. Konstan, and John Riedl. 2000. Analysis of recommendation algorithms for e-commerce. InACM Conference on Economics and Computation. https://api.semanticscholar.org/CorpusID:12366165

    2000

  33. [41]

    Tobias Schnabel and Paul N Bennett. 2020. Debiasing item-to-item recommenda- tions with small annotated datasets. InProceedings of the 14th ACM Conference on Recommender Systems. 73–81

    2020

  34. [42]

    Xiaoyuan Su and Taghi M Khoshgoftaar. 2009. A survey of collaborative filtering techniques.Advances in Artificial Intelligence2009 (2009)

    2009

  35. [43]

    Qiaoyu Tan, Jianwei Zhang, Jiangchao Yao, Ninghao Liu, Jingren Zhou, Hongxia Yang, and Xia Hu. 2021. Sparse-interest network for sequential recommendation. InProceedings of the 14th ACM International Conference on Web Search and Data Mining. 598–606

    2021

  36. [44]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971(2023)

    2023 arXiv

  37. [45]

    Hangyu Wang, Jianghao Lin, Xiangyang Li, Bo Chen, Chenxu Zhu, Ruiming Tang, Weinan Zhang, and Yong Yu. 2023. FLIP: Fine-grained Alignment between ID-based Models and Pretrained Language Models for CTR Prediction. InACM Conference on Recommender Systems. https://api.semanticsch...

    2023

  38. [46]

    Kini, Devendra Yadav, Fei Wang, Zhen Wen, Jiliang Tang, and Hui Liu

    Hanbing Wang, Xiaorui Liu, Wenqi Fan, Xiangyu Zhao, Venkataramana B. Kini, Devendra Yadav, Fei Wang, Zhen Wen, Jiliang Tang, and Hui Liu. 2024. Rethinking Large Language Model Architectures for Sequential Recommendations.ArXiv abs/2402.09543 (2024). https://api.semanticscholar...

    2024 arXiv

  39. [47]

    Qi Wang, Jindong Li, Shiqi Wang, Qianli Xing, Runliang Niu, He Kong, Rui Li, Guodong Long, Yi Chang, and Chengqi Zhang. 2024. Towards Next-Generation LLM-based Recommender Systems: A Survey and Beyond.ArXivabs/2410.19744 (2024). https://api.semanticscholar.org/CorpusID:273653859

    2024 arXiv

  40. [48]

    Yancheng Wang, Ziyan Jiang, Zheng Chen, Fan Yang, Yingxue Zhou, Eunah Cho, Xing Fan, Xiaojiang Huang, Yanbin Lu, and Yingzhen Yang. 2023. Rec- mind: Large language model powered agent for recommendation.arXiv preprint arXiv:2308.14296(2023)

    2023 arXiv

  41. [49]

    Yuyan Wang, Mohit Sharma, Can Xu, Sriraj Badam, Qian Sun, Lee Richardson, Lisa Chung, Ed H Chi, and Minmin Chen. 2022. Surrogate for long-term user experience in recommender systems. InProceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining. 4100–4109

    2022

  42. [50]

    Wei Wei, Xubin Ren, Jiabin Tang, Qinyong Wang, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. 2023. LLMRec: Large Language Models with Graph Augmentation for Recommendation.Proceedings of the 17th ACM International Conference on Web Search and Data Mining(2023)...

    2023

  43. [51]

    Likang Wu, Zhilan Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, Hui Xiong, and Enhong Chen. 2023. A Survey on Large Language Models for Recommendation.ArXiv abs/2305.19860 (2023). https://api.semanticscholar.org/CorpusID:25...

    2023 arXiv

  44. [52]

    Yunjia Xi, Weiwen Liu, Jianghao Lin, Jieming Zhu, Bo Chen, Ruiming Tang, Weinan Zhang, Rui Zhang, and Yong Yu. 2023. Towards Open-World Recom- mendation with Knowledge Augmentation from Large Language Models.ArXiv abs/2306.10933 (2023). https://api.semanticscholar.org/CorpusID...

    2023 arXiv

  45. [53]

    Zhibo Xiao, Luwei Yang, Tao Zhang, Wen Jiang, Wei Ning, and Yujiu Yang. 2024. Deep Evolutional Instant Interest Network for CTR Prediction in Trigger-Induced Recommendation.ArXiv(2024)

    2024

  46. [54]

    Xin Xin, Fajie Yuan, Xiangnan He, and Joemon M Jose. 2018. Batch is not heavy: Learning word representations from all samples. InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics. 1853–1862

    2018

  47. [55]

    Hu Xu, Saining Xie, Xiaoqing Ellen Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichten- hofer. 2023. Demystifying clip data.arXiv preprint arXiv:2309.16671(2023)

    2023 arXiv

  48. [56]

    Xiaohan Xu, Ming Li, Chongyang Tao, Tao Shen, Reynold Cheng, Jinyang Li, Can Xu, Dacheng Tao, and Tianyi Zhou. 2024. A Survey on Knowledge Distillation of Large Language Models.ArXivabs/2402.13116 (2024). https://api.semanticscholar. org/CorpusID:267760021

    2024 arXiv

  49. [57]

    Zhenhui Xu, Meng Zhao, Liqun Liu, Lei Xiao, Xiaopeng Zhang, and Bifeng Zhang

  50. [58]

    Feng Xue, Xiangnan He, Xiang Wang, Jiandong Xu, Kai Liu, and Richang Hong

  51. [59]

    Jing Yan, Liu Jiang, Jianfei Cui, Zhichen Zhao, Xingyan Bin, Feng Zhang, and Zuotao Liu. 2024. Trinity: Syncretizing Multi-/Long-Tail/Long-Term Interests All in One. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Sp...

    2024

  52. [60]

    Ji Yang, Xinyang Yi, Derek Zhiyuan Cheng, Lichan Hong, Yang Li, Simon Xiaom- ing Wang, Taibai Xu, and Ed H Chi. 2020. Mixed negative sampling for learning two-tower neural networks in recommendations. InCompanion proceedings of the web conference 2020. 441–447

    2020

  53. [61]

    Yantao Yu, Weipeng Wang, Zhoutian Feng, and Daiyue Xue. 2021. A dual aug- mented two-tower model for online large-scale recommendation. InDLP-KDD

    2021

  54. [62]

    Jiaqi Zhai, Zhaojie Gong, Yueming Wang, Xiao Sun, Zheng Yan, Fu Li, and Xing Liu. 2023. Revisiting Neural Retrieval on Accelerators. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 5520–5531

    2023

  55. [63]

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhao- jie Gong, Fangda Gu, Michael He, et al. 2024. Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations.arXiv preprint arXiv:2402.17152(2024)

    2024 arXiv

  56. [64]

    Weizhi Zhang, Yuan-Qi Bei, Liangwei Yang, Henry Peng Zou, Peilin Zhou, Aiwei Liu, Yinghui Li, Hao Chen, Jianling Wang, Yu Wang, Feiran Huang, Sheng Zhou, Jiajun Bu, Allen Lin, James Caverlee, Fakhri Karray, Irwin King, and Philip S. Yu

  57. [65]

    Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jianyun Nie, and Ji rong Wen

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Z. Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jianyun Nie, and Ji rong Wen. 20...

    2023 arXiv

  58. [66]

    Zhi Zheng, WenShuo Chao, Zhaopeng Qiu, Hengshu Zhu, and Hui Xiong

  59. [72]

    https: //api.semanticscholar.org/CorpusID:268536921

    Harnessing Large Language Models for Text-Rich Sequential Recom- mendation.Proceedings of the ACM on Web Conference 2024(2024). https: //api.semanticscholar.org/CorpusID:268536921

    2024

  60. [2012]

    InInternational Conference on Machine Learning

    BPR: Bayesian personalized ranking from implicit feedback. InInternational Conference on Machine Learning. 452–460

  61. [2019]

    Deep item-based collaborative filtering for top-n recommendation.ACM Transactions on Information Systems (TOIS)37, 3 (2019), 1–25

    2019

  62. [2022]

    InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Mixture of virtual-kernel experts for multi-objective user profile modeling. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4257–4267

  63. [2024]

    https://api.semanticscholar

    Large Language Model Enhanced Recommender Systems: Taxonomy, Trend, Application and Future.ArXivabs/2412.13432 (2024). https://api.semanticscholar. org/CorpusID:274822665

    2024 arXiv

  64. [2025]

    https://api.semanticscholar

    Cold-Start Recommendation towards the Era of Large Language Models (LLMs): A Comprehensive Survey and Roadmap. https://api.semanticscholar. org/CorpusID:275323883

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.