Pith. sign in

REVIEW 3 major objections 5 minor 74 references

Listwise Preference Alignment Optimization for Tail Item Recommendation

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that replacing pairwise preference comparison with a listwise softmax loss, plus adaptive negative sampling and reweighting, improves tail-item recommendation by up to 50% at lower training cost than DPO.

desk verdict A plausible empirical recipe for tail-item recommendation, but the theoretical claim does not follow and Table I contradicts the headline. read the letter →

arxiv 2507.02255 v1 pith:WBEPMAXH submitted 2025-07-03 cs.IR cs.LG

Zihao Li , Chao Yang , Tong Zhang , Yakun Chen , Xianzhi Wang , Guandong Xu , Daoyi Dong This is my paper
classification cs.IRcs.LG
keywords listwisepreferenceoptimizationtailitemrecommendationsequentialdirectBradley-TerrymodeladaptivenegativesamplingreweightingGumbel-Softmax
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

Preference alignment usually compares one accepted item to one rejected item, or requires a separately trained reward model. This paper tries to make the comparison listwise: one positive item versus K negatives in a single softmax, with the reference model removed, so each training step learns from many negatives at lower compute. The authors claim that optimizing this listwise preference optimization (LPO) loss is equivalent to maximizing an upper bound of the optimal reward under a listwise Bradley-Terry model. They pair the loss with adaptive sampling of negatives from the head items and reweighting of batches toward tail items, and report up to 50% improvement in tail-item metrics on three Amazon datasets while using less GPU memory than direct preference optimization (DPO). If the claim holds, a cheap training loss can redirect sequential recommenders toward the long tail without an explicit reward model.

What carries the argument

The load-bearing object is the listwise preference optimization loss, a softmax over one positive item and K negatives in which the logits are the model's own raw prediction probabilities scaled by a temperature $\tau$. It is derived by multiplying pairwise Bradley-Terry win probabilities, bounding the product with the inequality $\prod_i(1+x_i)>1+\sum_i x_i$ to get a tractable upper bound, and then following SimPO and ORPO in dropping the reference model and exponentiating the probability ratios. Around this loss the paper wraps two mechanisms: adaptive negative sampling, which selects negatives among head items with Gumbel-Softmax according to their current predicted probabilities, and adaptive reweighting, which assigns higher weight to samples whose ground-truth item is a tail item.

What would settle it

Run LPO on data where an independent reward model supplies ground-truth preference scores and check whether the items LPO ranks highest are the ones that reward model scores highest; if minimizing the LPO loss does not push the model toward that reward ranking on held-out items, the claimed equivalence with optimal-reward maximization is not supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that preference alignment can be performed with a listwise Bradley-Terry comparison whose optimal reward has a closed form, so neither a reward model nor a reference policy is needed. The paper derives the LPO loss $L = -\log \frac{\exp(\pi_\theta(y_w)/\tau)}{\exp(\pi_\theta(y_w)/\tau)+\sum_{\ell=1}^K \exp(\pi_\theta(y_\ell)/\tau)}$ and argues that minimizing it maximizes an upper bound of the optimal policy reward. It further claims that drawing negatives from the head items via Gumbel-Softmax and reweighting toward tail positives steers optimization toward tail items, yielding up to 50% gains in tail-item Hit Rate and NDCG with lower memory and training cost than DPO and S-DPO.

Load-bearing premise

The load-bearing premise is that the model's raw probability scores behave like preference rewards once the reference model is discarded, so exponentials of those scores can fill the role of reward logits.

Editorial extensions

If this is right

  • Training can exploit K negatives per positive in a single softmax step, so negative examples are used more thoroughly than in pairwise DPO.
  • The reference model can be dropped entirely, cutting peak GPU memory to 82.1% of DPO's and removing a whole model copy from training.
  • The LPO loss is model-agnostic: grafting it onto SASRec, GRU4Rec, or Caser improves their tail-item metrics, not just the proposed pipeline.
  • Because negatives are drawn from the head, the model's probability mass shifts toward tail items, which the paper shows as a rightward shift in the distribution of tail-item recommendation probabilities.
  • Hyperparameter trends indicate a controllable trade-off: larger tail-reweighting weight $\alpha_T$ and larger LPO weight $\lambda$ improve tail metrics while modestly lowering overall metrics.

Reading between the lines

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

  • My inference: the LPO loss should transfer to non-sequential recommenders and retrieval models, since it only requires one positive score and K negative scores from the same model.
  • My inference: if the raw-probability-as-reward assumption is miscalibrated, the loss still works as a ranking objective, but the claimed equivalence to optimal-reward maximization would require recalibration; comparing LPO rankings against an explicit reward model on held-out items would settle this.
  • My inference: because negatives are deliberately drawn from the head, the recipe should also raise catalog coverage and diversity metrics, not just tail-item hit rate; this could be checked directly on the same datasets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. This paper proposes LPO4Rec, a training objective for sequential recommendation that extends the pairwise Bradley-Terry comparison in DPO to a listwise comparison against K negative items, removes the reference model, and combines the resulting softmax-style loss with a cross-entropy loss, adaptive negative sampling from head items, and a reweighting scheme that up-weights tail-item ground truths. The authors claim that optimizing the LPO loss is equivalent to maximizing an upper bound of the optimal policy under a listwise Bradley-Terry model, and report consistent gains over ten baselines on three Amazon datasets, with up to 50% improvement on tail-item metrics and lower GPU memory than DPO/S-DPO. The paper also reports ablations, hyperparameter analyses, and efficiency measurements.

Significance. If the theoretical equivalence were valid, the paper would offer a simple, model-agnostic and memory-efficient recipe for tail-item recommendation, with a useful extension of preference alignment beyond pairwise comparisons. The empirical study is broad: ten baselines, three datasets, all-item candidate evaluation, ablations, and per-epoch time/memory measurements, and the authors release code. However, the central derivation in Section IV-A is mathematically invalid, and the main empirical table contains an internal contradiction on the Amazon Sports tail rows. Because the claimed theory is the paper's headline contribution and the empirical inconsistency concerns the headline 50% improvement, the significance of the work is not currently established. The practical recipe may still be useful, but the manuscript as written does not support its stated contributions.

major comments (3)
  1. [IV-A, Eq. (11)] The step marked '∝' in Eq. (11) is not a proportionality. Starting from Eq. (10), each factor is a/(a+b) with a=π*_θ(yw|x)/πref(yw|x) and b=π*_θ(yℓ|x)/πref(yℓ|x); the paper replaces it with exp(a)/(exp(a)+exp(b)). The ratio of the two expressions depends on b: for a=0.2, b=0.4 it is approximately 1.040, while for a=0.2, b=0.8 it is approximately 1.218, so no constant rescaling exists. Consequently the inequality ∏(1+x_i)>1+Σx_i is applied after a change of functional form, and the claimed upper bound on the original listwise Bradley-Terry probability in Eq. (10) is not established. Moreover, if q denotes the softmax expression and p the true listwise BT probability, the relationship p<q would imply -log q < -log p, so minimizing the LPO loss does not minimize the negative log-likelihood of the true objective. The central theoretical claim that optimizing LPO is equivalent to maximizing an upper bound of the optimal policy/reward is therefore unsupported.
  2. [Table I, Amazon Sports tail rows] The claim in Section V-C that 'LPO4Rec consistently outperforms the baselines across all metrics' is contradicted by Table I. In the Amazon Sports tail-item rows, GRU4Rec achieves HR@5=0.0066 and HR@20=0.0119, while LPO4Rec achieves 0.0054 and 0.0096, respectively; the reported ▲%=50.00% for tail HR@20 appears to be computed against R2Rec (0.0064) rather than against the best of all listed baselines. This internal inconsistency undermines the headline 'up to 50% improvement' claim and requires correction before the empirical contribution can be assessed.
  3. [IV-D, Eq. (19)] The gradient derivation in Eqs. (18)-(19) does not constitute an independent theoretical proof of hard-item focus and output diversity. The 'hard-item' property follows directly from the algebraic form of the softmax loss, and the 'tail-item/diversity' property relies on the fact that negative samples are drawn from the head set by construction in Eq. (14) and reweighted by Eq. (16). The informal approximation exp(πθ(y|x)) ≪ K E[exp(πθ(yℓ|x))] and the subsequent '∝' step are not justified, and the claims in Section II-C and the contributions list that LPO 'theoretically proves' these properties are overstated.
minor comments (5)
  1. [IV-A, Eq. (12)] The equality between the first and second displayed forms is not valid when τ≠1; the first form omits the temperature in the exponent. Introduce τ consistently in both expressions or state that the first form corresponds to τ=1.
  2. [IV-C, Eq. (16)] The text 'αT > αT' should read 'αT > αH'.
  3. [IV-C, Eq. (15)] The Gumbel noise notation is garbled; 'Gumbel(0, β) ∼ −βlog(−log(0, β))' should be '−log(−log U)' with U∼Uniform(0,1), and the text should distinguish Gumbel-Softmax relaxation from Gumbel-Top-K sampling.
  4. [Table I, Amazon Sports overall HR@20] The ORPO entry 0.0024 is an outlier compared to all other ORPO entries (about 0.06) and is likely a typographical error; please verify.
  5. [IV-A, Eq. (9)] The product of pairwise Bradley-Terry probabilities is not the same as a Plackett-Luce listwise model; the term 'listwise Bradley-Terry' should be defined precisely to avoid confusion with the softmax/PL model used by S-DPO.

Circularity Check

2 steps flagged · score 5.0 of 10

Central 'upper-bound equivalence' is a self-constructed loss identity; tail-diversity merit restates the adaptive sampling/reweighting design.

  1. self definitional [Section IV-A, Eqs. (11)-(12)]
    "Further, considering QK i=1(1 +xi) > 1 +PK i=1 xi when xi > 0, we re-formalize Eq. (10) into: ... ∝ ... < 1/(1+PK ℓ=1 exp(...)). ... Finally, the LPO loss is derived from the maximum likelihood objective of Eq. (11), as formalized below: LLPO(πθ) = −E log 1/(1+PK ℓ=1 exp(πθ(yℓ|x) − πθ(yw|x))). Comparing Eq. (12) with Eq. (11), we optimize LLPO to closely approximate the upper bounds of the optimal policy model π∗ref within the Bradley-Terry framework."

    The 'upper bound' in Eq. (11) is manufactured by replacing each true listwise Bradley-Terry factor a/(a+b) with exp(a)/(exp(a)+exp(b)); no proportionality constant exists, so this is a change of functional form, not a bound on Eq. (9). Eq. (12) is then deliberately defined as minus the log of that manufactured expression. Consequently, the claim that optimizing LPO maximizes the upper bound of the optimal policy is an identity between the loss and the expression used to construct it, not a derived connection to the Bradley-Terry optimal reward. The central theoretical assertion therefore reduces to the loss's own definition plus an algebraic error.

  2. self definitional [Section IV-C, Eq. (16); Section IV-D, after Eq. (19)]
    "Here, we define αT = 1, αH = 0, ensuring αT > αT—this allows the model to assign greater weight to samples from the tail item set during optimization. ... According to the denominator of Eq. (19), negative samples yℓ are predominantly drawn from the head items, given that sampling probability is proportional to the recommendation probability ... Accordingly, the recommended positive items (e.g., yw) have a preference for tail items, promoting diversity and recommendations of tail items."

    The tail-diversity 'merit' is not an emergent theoretical result: Eq. (14) already restricts negative sampling to the head-item set IH, and Eq. (16) directly up-weights tail-item training samples via αT = 1, αH = 0. The gradient reasoning in Eq. (19) merely reflects these built-in design choices. Presenting this as a theoretical property of LPO is a restatement of the loss's own sampling and reweighting construction, so the claimed tail-item enhancement is by construction rather than independently derived.

full rationale

The paper's empirical comparisons on three public datasets are external and not circular: the reported tail-item gains are concrete outcomes of a specific training recipe, and the reweighting strategy is additionally benchmarked against the authors' earlier R2Rec baseline. However, the central theoretical claim is self-referential in structure. Starting from Eq. (10), the paper replaces each listwise BT factor a/(a+b) by exp(a)/(exp(a)+exp(b)) with a '∝' that no constant can justify; after this algebraic change, it builds an upper bound and then defines the LPO loss as the negative log of that upper bound. Hence the statement that minimizing LPO maximizes the upper bound is true by construction—it is an identity between the loss and the expression crafted for it—and cannot independently validate the connection to the Bradley-Terry optimal policy. The algebraic invalidity of the proportionality is fundamentally a correctness risk, but the circular aspect is that the 'proof' validates the loss by reference to an expression deliberately built to be that loss. Separately, the claimed merits of hard-item focus and tail-item diversity are gradient and sampling properties of the loss, the adaptive head-item negative sampling, and the tail-item reweighting; these are restatements of the design rather than first-principles predictions. The paper is therefore partially circular in its theoretical narrative, while its empirical content remains independently testable.

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

The loss has four hand-set hyperparameters that directly encode tail-item priority, and the 'derivation' from Bradley-Terry relies on an unproven exponential substitution. No new entities are introduced.

free parameters (5)
  • λ (LPO loss weight) = 0.5 in main experiments; recommended 0.3 after hyperparameter study
    Balances cross-entropy and LPO loss; chosen by hand and tuned.
  • τ (temperature in LPO softmax) = 0.1 in main experiments; range [0.05,2] explored
    Controls sharpness of the preference distribution; hand-set.
  • K (number of negative samples) = Not stated for main experiments; tuned over [2,50], recommended 10
    Number of rejected items per positive; affects sampling distribution and loss shape.
  • αT and αH (reweighting exponents) = αT=1, αH=0
    Directly encodes tail priority in the optimization objective.
  • Head/tail split ratio = Pareto 80/20
    Determines which items are labeled tail and thus the reported tail metrics.
assumptions (5)
  • standard math Bradley-Terry model and DPO closed-form optimal reward
    Borrowed from Rafailov et al.; used in Eqs. (6)-(8).
  • ad hoc to paper Model probabilities can act as reward logits without a reference model
    Dropping πref and exponentiating in Eq. (11)-(12) is assumed, not derived.
  • ad hoc to paper The '∝' replacement in Eq. (11) preserves the optimization semantics
    Mathematically, a/b is not proportional to e^a/(e^a+e^b); this step is undefined.
  • domain assumption Law of large numbers / large-K approximation in gradient analysis
    Eq. (19) replaces a finite sum with K times an expectation and assumes exp(π_w) is negligible.
  • domain assumption Popular head items are a useful negative distribution for improving tail items
    The adaptive sampler draws only from IH; if false, the method could suppress head items without lifting tail items.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Listwise Preference Alignment Optimization for Tail Item Recommendation." pith.science (2026). https://pith.science/paper/WBEPMAXH

@misc{pith2026250702255,
  author       = {Pith},
  title        = {Pith review of: Listwise Preference Alignment Optimization for Tail Item Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WBEPMAXH}},
  note         = {Machine review of arXiv:2507.02255}
}
read the original abstract

Preference alignment has achieved greater success on Large Language Models (LLMs) and drawn broad interest in recommendation research. Existing preference alignment methods for recommendation either require explicit reward modeling or only support pairwise preference comparison. The former directly increases substantial computational costs, while the latter hinders training efficiency on negative samples. Moreover, no existing effort has explored preference alignment solutions for tail-item recommendation. To bridge the above gaps, we propose LPO4Rec, which extends the Bradley-Terry model from pairwise comparison to listwise comparison, to improve the efficiency of model training. Specifically, we derive a closed form optimal policy to enable more efficient and effective training without explicit reward modeling. We also present an adaptive negative sampling and reweighting strategy to prioritize tail items during optimization and enhance performance in tail-item recommendations. Besides, we theoretically prove that optimizing the listwise preference optimization (LPO) loss is equivalent to maximizing the upper bound of the optimal reward. Our experiments on three public datasets show that our method outperforms 10 baselines by a large margin, achieving up to 50% performance improvement while reducing 17.9% GPU memory usage when compared with direct preference optimization (DPO) in tail-item recommendation. Our code is available at https://github.com/Yuhanleeee/LPO4Rec.

Figures

Figures reproduced from arXiv: 2507.02255 by the authors.

Figure 1
Figure 1. Performance of SASRec and SASRec with our LPO loss function [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison of DPO (left) and LPO (right) loss functions. DPO optimizes the model with pairwise ( [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The single-pass (one epoch) training time and peak GPU memory [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The overall and tail item performance under different hyperparameter settings. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: The frequency distribution of the differences between the average [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

74 extracted references · 55 canonical work pages

  1. [1]

    Harnessing the power of llms in practice: A survey on chatgpt and beyond,

    J. Yang, H. Jin, R. Tang, X. Han, Q. Feng, H. Jiang, S. Zhong, B. Yin, and X. Hu, “Harnessing the power of llms in practice: A survey on chatgpt and beyond,” ACM Transactions on Knowledge Discovery from Data, vol. 18, no. 6, pp. 1–32, 2024

    2024

  2. [2]

    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, and Q. Li, “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

    work page 2024

  3. [3]

    Tallrec: An effective and efficient tuning framework to align large language model with recommendation,

    K. Bao, J. Zhang, Y . Zhang, W. Wang, F. Feng, and X. He, “Tallrec: An effective and efficient tuning framework to align large language model with recommendation,” in Proceedings of the 17th ACM Conference on Recommender Systems, 2023, pp. 1007–1014

    work page 2023

  4. [4]

    Aligning large language model with direct multi-preference optimization for recommendation,

    Z. Bai, N. Wu, F. Cai, X. Zhu, and Y . Xiong, “Aligning large language model with direct multi-preference optimization for recommendation,” in Proceedings of the 33rd ACM International Conference on Information and Knowledge Management , 2024, pp. 76–86

    work page 2024

  5. [5]

    Who to align with: Feedback-oriented multi-modal alignment in recommendation systems,

    Y . Li, Q. Zhao, C. Lin, J. Su, and Z. Zhang, “Who to align with: Feedback-oriented multi-modal alignment in recommendation systems,” in Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2024, pp. 667– 676

    work page 2024

  6. [6]

    Onerec: Unifying retrieve and rank with generative recommender and iterative preference alignment,

    J. Deng, S. Wang, K. Cai, L. Ren, Q. Hu, W. Ding, Q. Luo, and G. Zhou, “Onerec: Unifying retrieve and rank with generative recommender and iterative preference alignment,” arXiv preprint arXiv:2502.18965, 2025

    arXiv 2025

  7. [7]

    A survey on human preference learning for large language models,

    R. Jiang, K. Chen, X. Bai, Z. He, J. Li, M. Yang, T. Zhao, L. Nie, and M. Zhang, “A survey on human preference learning for large language models,” arXiv preprint arXiv:2406.11191 , 2024

    arXiv 2024

  8. [8]

    Training language models to follow instructions with human feedback,

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray et al. , “Training language models to follow instructions with human feedback,”Advances in Neural Information Processing Systems , vol. 35, pp. 27 730–27 744, 2022

    work page 2022

Show all 74 references
  1. [9]

    Leveraging large language models for sequential recommendation,

    J. Harte, W. Zorgdrager, P. Louridas, A. Katsifodimos, D. Jannach, and M. Fragkoulis, “Leveraging large language models for sequential recommendation,” in Proceedings of the 17th ACM Conference on Recommender Systems, 2023, pp. 1096–1102

    2023

  2. [10]

    Harnessing large language models for text-rich sequential recommendation,

    Z. Zheng, W. Chao, Z. Qiu, H. Zhu, and H. Xiong, “Harnessing large language models for text-rich sequential recommendation,” in Proceedings of the ACM Web Conference , 2024, pp. 3207–3216

    2024

  3. [11]

    Prox- imal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Prox- imal policy optimization algorithms,” arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  4. [12]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models,

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . Li, Y . Wuet al., “Deepseekmath: Pushing the limits of mathematical reasoning in open language models,” arXiv preprint arXiv:2402.03300 , 2024

    2024 arXiv

  5. [13]

    Dapo: An open-source llm reinforcement learning system at scale,

    Q. Yu, Z. Zhang, R. Zhu, Y . Yuan, X. Zuo, Y . Yue, T. Fan, G. Liu, L. Liu, X. Liu et al., “Dapo: An open-source llm reinforcement learning system at scale,” arXiv preprint arXiv:2503.14476 , 2025

    2025 arXiv

  6. [14]

    Direct preference optimization: Your language model is secretly a reward model,

    R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn, “Direct preference optimization: Your language model is secretly a reward model,” Advances in Neural Information Processing Systems, vol. 36, pp. 53 728–53 741, 2023

    2023

  7. [15]

    Simpo: Simple preference optimization with a reference-free reward,

    Y . Meng, M. Xia, and D. Chen, “Simpo: Simple preference optimization with a reference-free reward,” Advances in Neural Information Process- ing Systems, vol. 37, pp. 124 198–124 235, 2024

    2024

  8. [16]

    Reference-free monolithic preference optimization with odds ratio,

    J. Hong, N. Lee, and J. Thorne, “Reference-free monolithic preference optimization with odds ratio,” arXiv preprint arXiv:2403.07691v1, 2024

    2024 arXiv

  9. [17]

    Rank analysis of incomplete block designs: I. the method of paired comparisons,

    R. A. Bradley and M. E. Terry, “Rank analysis of incomplete block designs: I. the method of paired comparisons,” Biometrika, vol. 39, no. 3/4, pp. 324–345, 1952

    1952

  10. [18]

    On softmax direct preference optimization for recommen- dation,

    Y . Chen, J. Tan, A. Zhang, Z. Yang, L. Sheng, E. Zhang, X. Wang, and T.-S. Chua, “On softmax direct preference optimization for recommen- dation,” arXiv preprint arXiv:2406.09215 , 2024

    2024 arXiv

  11. [19]

    On the theories behind hard negative sampling for recommendation,

    W. Shi, J. Chen, F. Feng, J. Zhang, J. Wu, C. Gao, and X. He, “On the theories behind hard negative sampling for recommendation,” in Proceedings of the ACM Web Conference , 2023, pp. 812–822

    2023

  12. [20]

    Negative sampling in recommendation: A survey and future directions,

    H. Ma, R. Xie, L. Meng, F. Feng, X. Du, X. Sun, Z. Kang, and X. Meng, “Negative sampling in recommendation: A survey and future directions,” arXiv preprint arXiv:2409.07237 , 2024

    2024 arXiv

  13. [21]

    An analysis for unreplicated fractional factorials,

    G. E. Box and R. D. Meyer, “An analysis for unreplicated fractional factorials,” Technometrics, vol. 28, no. 1, pp. 11–18, 1986

    1986

  14. [22]

    Justifying recommendations using distantly-labeled reviews and fine-grained aspects,

    J. Ni, J. Li, and J. McAuley, “Justifying recommendations using distantly-labeled reviews and fine-grained aspects,” in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Pro- cessing and the 9th International Joint Conference on Natural Language Proce...

    2019

  15. [23]

    An mdp-based recom- mender system,

    G. Shani, D. Heckerman, and R. I. Brafman, “An mdp-based recom- mender system,” Journal of Machine Learning Research , vol. 6, no. Sep, pp. 1265–1295, 2005

    2005

  16. [24]

    Factorizing personalized markov chains for next-basket recommendation,

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

    2010

  17. [25]

    Deep learning based rec- ommender system: A survey and new perspectives,

    S. Zhang, L. Yao, A. Sun, and Y . Tay, “Deep learning based rec- ommender system: A survey and new perspectives,” ACM Computing Surveys, vol. 52, no. 1, pp. 1–38, 2019

    2019

  18. [26]

    Graph and sequential neural networks in session-based recommendation: A survey,

    Z. Li, C. Yang, Y . Chen, X. Wang, H. Chen, G. Xu, L. Yao, and M. Sheng, “Graph and sequential neural networks in session-based recommendation: A survey,” ACM Computing Surveys , vol. 57, no. 2, pp. 1–37, 2024

    2024

  19. [27]

    A survey on deep learning: Algorithms, techniques, and applications,

    S. Pouyanfar, S. Sadiq, Y . Yan, H. Tian, Y . Tao, M. P. Reyes, M.-L. Shyu, S.-C. Chen, and S. S. Iyengar, “A survey on deep learning: Algorithms, techniques, and applications,” ACM Computing Surveys , vol. 51, no. 5, pp. 1–36, 2018

    2018

  20. [28]

    Deep learning,

    Y . LeCun, Y . Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, 2015

    2015

  21. [29]

    Deep multimodal learning for time series analysis in social computing: a survey,

    C. Yang, Y . Chen, Z. Li, X. Wang, K. Shi, L. Yao, G. Xu, and Z. Guo, “Deep multimodal learning for time series analysis in social computing: a survey,” International Journal of Multimedia Information Retrieval , vol. 14, no. 2, p. 15, 2025

    2025

  22. [30]

    Personalized top-n sequential recommendation via convolutional sequence embedding,

    J. Tang and K. Wang, “Personalized top-n sequential recommendation via convolutional sequence embedding,” inProceedings of the 11th ACM International Conference on Web Search and Data Mining , 2018, pp. 565–573

    2018

  23. [31]

    Session- based recommendations with recurrent neural networks,

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

    2015 arXiv

  24. [32]

    Adversarial and contrastive variational autoencoder for sequential recommendation,

    Z. Xie, C. Liu, Y . Zhang, H. Lu, D. Wang, and Y . Ding, “Adversarial and contrastive variational autoencoder for sequential recommendation,” in Proceedings of the ACM Web Conference , 2021, pp. 449–459

    2021

  25. [33]

    Self-attentive sequential recommenda- tion,

    W.-C. Kang and J. McAuley, “Self-attentive sequential recommenda- tion,” in Proceedings of the 18th IEEE International Conference on Data Mining. IEEE, 2018, pp. 197–206

    2018

  26. [34]

    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 Con- ference on Information and Knowledge Management , 2019, pp. 1441– 1450

    2019

  27. [35]

    Continuous- time sequential recommendation with temporal graph collaborative transformer,

    Z. Fan, Z. Liu, J. Zhang, Y . Xiong, L. Zheng, and P. S. Yu, “Continuous- time sequential recommendation with temporal graph collaborative transformer,” in Proceedings of the 30th ACM International Conference on Information and Knowledge Management , 2021, pp. 433–442

    2021

  28. [36]

    Sse-pt: Sequential recommendation via personalized transformer,

    L. Wu, S. Li, C.-J. Hsieh, and J. Sharpnack, “Sse-pt: Sequential recommendation via personalized transformer,” in Proceedings of the 14th ACM Conference on Recommender Systems , 2020, pp. 328–337

    2020

  29. [37]

    Multi-behavior sequential recommendation with temporal graph transformer,

    L. Xia, C. Huang, Y . Xu, and J. Pei, “Multi-behavior sequential recommendation with temporal graph transformer,” IEEE Transactions on Knowledge and Data Engineering , vol. 35, no. 6, pp. 6099–6112, 2022

    2022

  30. [38]

    Kerl: A knowledge-guided reinforcement learning model for sequential rec- ommendation,

    P. Wang, Y . Fan, L. Xia, W. X. Zhao, S. Niu, and J. Huang, “Kerl: A knowledge-guided reinforcement learning model for sequential rec- ommendation,” in Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, 2020, pp. 209–218

    2020

  31. [39]

    Diffurec: A diffusion model for sequential recommendation,

    Z. Li, A. Sun, and C. Li, “Diffurec: A diffusion model for sequential recommendation,” ACM Transactions on Information Systems , vol. 42, no. 3, pp. 1–28, 2023

    2023

  32. [40]

    Diffusion recommender model,

    W. Wang, Y . Xu, F. Feng, X. Lin, X. He, and T.-S. Chua, “Diffusion recommender model,” in Proceedings of the 46th International ACM SI- GIR Conference on Research and Development in Information Retrieval, 2023, pp. 832–841

    2023

  33. [41]

    Towards universal sequence representation learning for recommender systems,

    Y . Hou, S. Mu, W. X. Zhao, Y . Li, B. Ding, and J.-R. Wen, “Towards universal sequence representation learning for recommender systems,” in Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2022, pp. 585–593

    2022

  34. [42]

    Text is all you need: Learning language representations for sequential recommendation,

    J. Li, M. Wang, J. Li, J. Fu, X. Shen, J. Shang, and J. McAuley, “Text is all you need: Learning language representations for sequential recommendation,” in Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2023, pp. 1258–1267

    2023

  35. [43]

    Recommendation as language processing (rlp): A unified pretrain, personalized prompt & JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2025 12 predict paradigm (p5),

    S. Geng, S. Liu, Z. Fu, Y . Ge, and Y . Zhang, “Recommendation as language processing (rlp): A unified pretrain, personalized prompt & JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2025 12 predict paradigm (p5),” in Proceedings of the 16th ACM Conference on Recommender ...

    2025

  36. [44]

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

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

    2024

  37. [45]

    Towards efficient and effective adaptation of large language models for sequential recommendation,

    B. Peng, B. Burns, Z. Chen, S. Parthasarathy, and X. Ning, “Towards efficient and effective adaptation of large language models for sequential recommendation,” arXiv preprint arXiv:2310.01612 , 2023

    2023 arXiv

  38. [46]

    Raserec: Retrieval-augmented sequential recommendation,

    X. Zhao, B. Hu, Y . Zhong, S. Huang, Z. Zheng, M. Wang, H. Wang, and M. Zhang, “Raserec: Retrieval-augmented sequential recommendation,” arXiv preprint arXiv:2412.18378 , 2024

    2024 arXiv

  39. [47]

    Harnessing multimodal large language models for multimodal sequential recommendation,

    Y . Ye, Z. Zheng, Y . Shen, T. Wang, H. Zhang, P. Zhu, R. Yu, K. Zhang, and H. Xiong, “Harnessing multimodal large language models for multimodal sequential recommendation,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 39, no. 12, 2025, pp. 13 069– 13 077

    2025

  40. [48]

    Challenging the long tail recommendation,

    H. Yin, B. Cui, J. Li, J. Yao, and C. Chen, “Challenging the long tail recommendation,” Proceedings of the VLDB Endowment , vol. 5, no. 9, 2012

    2012

  41. [49]

    Two birds one stone: on both cold-start and long-tail recommendation,

    J. Li, K. Lu, Z. Huang, and H. T. Shen, “Two birds one stone: on both cold-start and long-tail recommendation,” in Proceedings of the 25th ACM International Conference on Multimedia , 2017, pp. 898–906

    2017

  42. [50]

    Correcting sample selection bias by unlabeled data,

    J. Huang, A. Gretton, K. Borgwardt, B. Sch ¨olkopf, and A. Smola, “Correcting sample selection bias by unlabeled data,” Advances in Neural Information Processing Systems , vol. 19, 2006

    2006

  43. [51]

    Reembedding and reweighting are needed for tail item sequential recommendation,

    Z. Li, Y . Chen, T. Zhang, and X. Wang, “Reembedding and reweighting are needed for tail item sequential recommendation,” in Proceedings of the ACM Web Conference , 2025, pp. 4925–4936

    2025

  44. [52]

    A model of two tales: Dual transfer learning framework for improved long-tail item recommendation,

    Y . Zhang, D. Z. Cheng, T. Yao, X. Yi, L. Hong, and E. H. Chi, “A model of two tales: Dual transfer learning framework for improved long-tail item recommendation,” in Proceedings of the ACM Web Conference , 2021, pp. 2220–2231

    2021

  45. [53]

    Co-occurrence embedding enhancement for long-tail problem in multi-interest recommendation,

    Y . Liu, X. Zhang, M. Zou, and Z. Feng, “Co-occurrence embedding enhancement for long-tail problem in multi-interest recommendation,” in Proceedings of the 17th ACM Conference on Recommender Systems , 2023, pp. 820–825

    2023

  46. [54]

    Melt: Mutual enhancement of long-tailed user and item for sequential recommendation,

    K. Kim, D. Hyun, S. Yun, and C. Park, “Melt: Mutual enhancement of long-tailed user and item for sequential recommendation,” in Proceed- ings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2023, pp. 68–77

    2023

  47. [55]

    Cities: Contextual inference of tail-item embeddings for sequential recommendation,

    S. Jang, H. Lee, H. Cho, and S. Chung, “Cities: Contextual inference of tail-item embeddings for sequential recommendation,” in Proceedings of the IEEE International Conference on Data Mining . IEEE, 2020, pp. 202–211

    2020

  48. [56]

    Metaga: Metalearning with graph-attention for improved long-tail item recommendation,

    B. Qin, Z. Huang, Z. Wu, C. Wang, and Y . Chen, “Metaga: Metalearning with graph-attention for improved long-tail item recommendation,”IEEE Transactions on Computational Social Systems , 2024

    2024

  49. [57]

    Loam: Improving long- tail session-based recommendation via niche walk augmentation and tail session mixup,

    H. Yang, Y . Choi, G. Kim, and J.-H. Lee, “Loam: Improving long- tail session-based recommendation via niche walk augmentation and tail session mixup,” in Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2023, pp...

    2023

  50. [58]

    Multifdf: Multi-community clustering for fairness-aware recommendation,

    B. Wang, S. Song, S. Liu, and X. Deng, “Multifdf: Multi-community clustering for fairness-aware recommendation,” IEEE Transactions on Computational Social Systems , vol. 10, no. 6, pp. 2959–2970, 2023

    2023

  51. [59]

    Memory bank augmented long-tail sequential recommendation,

    Y . Hu, Y . Liu, C. Miao, and Y . Miao, “Memory bank augmented long-tail sequential recommendation,” in Proceedings of the 31st ACM International Conference on Information and Knowledge Management , 2022, pp. 791–801

    2022

  52. [60]

    Large language models enhanced sequential recommendation for long- tail user and item,

    Q. Liu, X. Wu, X. Zhao, Y . Wang, Z. Zhang, F. Tian, and Y . Zheng, “Large language models enhanced sequential recommendation for long- tail user and item,” arXiv preprint arXiv:2405.20646 , 2024

    2024 arXiv

  53. [61]

    Multi- modal mixture of experts representation learning for sequential recom- mendation,

    S. Bian, X. Pan, W. X. Zhao, J. Wang, C. Wang, and J.-R. Wen, “Multi- modal mixture of experts representation learning for sequential recom- mendation,” in Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , 2023, pp. 110–119

    2023

  54. [62]

    Relation pruning and discriminative sampling over knowledge graph for long-tail recommendation,

    Z. Zhang, A. Wang, Y . Zhang, Y . Ren, W. Li, B. Wang, and M. Inuiguchi, “Relation pruning and discriminative sampling over knowledge graph for long-tail recommendation,” Information Sciences, p. 120871, 2024

    2024

  55. [63]

    Constitutional ai: Harmlessness from ai feedback,

    Y . Bai, S. Kadavath, S. Kundu, A. Askell, J. Kernion, A. Jones, A. Chen, A. Goldie, A. Mirhoseini, C. McKinnon et al., “Constitutional ai: Harmlessness from ai feedback,” arXiv preprint arXiv:2212.08073 , 2022

    2022 arXiv

  56. [64]

    Controllable preference optimization: toward controllable multi-objective alignment,

    Y . Guo, G. Cui, L. Yuan, N. Ding, Z. Sun, B. Sun, H. Chen, R. Xie, J. Zhou, Y . Lin et al. , “Controllable preference optimization: toward controllable multi-objective alignment,” in Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , 2024...

    2024

  57. [65]

    Arithmetic control of llms for diverse user prefer- ences: directional preference alignment with multi-objective rewards,

    H. Wang, Y . Lin, W. Xiong, R. Yang, S. Diao, S. Qiu, H. Zhao, and T. Zhang, “Arithmetic control of llms for diverse user prefer- ences: directional preference alignment with multi-objective rewards,” in Proceedings of the 62nd Annual Meeting of the Association for Computation...

    2024

  58. [66]

    Sequence tutor: Conservative fine-tuning of sequence generation models with kl-control,

    N. Jaques, S. Gu, D. Bahdanau, J. M. Hern ´andez-Lobato, R. E. Turner, and D. Eck, “Sequence tutor: Conservative fine-tuning of sequence generation models with kl-control,” in Proceedings of the International Conference on Machine Learning . PMLR, 2017, pp. 1645–1654

    2017

  59. [67]

    Direct preference optimization: Your language model is secretly a reward model,

    R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn, “Direct preference optimization: Your language model is secretly a reward model,” Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  60. [68]

    Relative entropy policy search,

    J. Peters, K. Mulling, and Y . Altun, “Relative entropy policy search,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 24, no. 1, 2010, pp. 1607–1612

    2010

  61. [69]

    Advantage-weighted regression: Simple and scalable off-policy reinforcement learning,

    X. B. Peng, A. Kumar, G. Zhang, and S. Levine, “Advantage-weighted regression: Simple and scalable off-policy reinforcement learning,”arXiv preprint arXiv:1910.00177, 2019

    1910 arXiv

  62. [70]

    The concrete distribution: A continuous relaxation of discrete random variables,

    C. J. Maddison, A. Mnih, and Y . W. Teh, “The concrete distribution: A continuous relaxation of discrete random variables,” arXiv preprint arXiv:1611.00712, 2016

    2016 arXiv

  63. [71]

    Stochastic beams and where to find them: The gumbel-top-k trick for sampling sequences without re- placement,

    W. Kool, H. Van Hoof, and M. Welling, “Stochastic beams and where to find them: The gumbel-top-k trick for sampling sequences without re- placement,” in Proceedings of the International Conference on Machine Learning. PMLR, 2019, pp. 3499–3508

    2019

  64. [72]

    On sampled metrics for item recommen- dation,

    W. Krichene and S. Rendle, “On sampled metrics for item recommen- dation,” Communications of the ACM , vol. 65, no. 7, pp. 75–83, 2022

    2022

  65. [73]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in Proceedings of the International Conference on Machine Learning . PMLR, 2021, pp. 8748–8763

    2021

  66. [74]

    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, pp. 103–111

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.