Pith. sign in

REVIEW 3 major objections 5 minor 67 references

Enhancing Temporal Sensitivity of Large Language Model for Recommendation with Counterfactual Tuning

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 Item-level temporal embeddings plus a counterfactual tuning objective make an LLM recommender order-aware and beat prior baselines across three real-world datasets.

desk verdict CETRec is a solid incremental method for LLM-based sequential recommendation, but the counterfactual loss in Eq. (8) is underspecified and must be clarified before the reported gains can be trusted. read the letter →

arxiv 2507.03047 v2 pith:UHGYJBG4 submitted 2025-07-03 cs.CL cs.AIcs.IR

Yutian Liu , Zhengyi Yang , Jiancan Wu , Xiang Wang This is my paper
classification cs.CLcs.AIcs.IR
keywords SequentialRecommendationLLM-basedCounterfactualInferenceTemporalEmbeddingSensitivityPositionCausal
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 LLM-based sequential recommenders underuse the temporal order of user interactions because Transformer self-attention and NLP-style position embeddings are weak at encoding interaction order. It proposes CETRec, which adds an item-level temporal embedding to every token of each item and trains with a counterfactual tuning loss that contrasts the ordered history against an order-erased version. The reported experiments on MovieLens, Steam, and LastFM show that the RoPE-based CETRec outperforms all compared traditional and LLM-based recommenders on Hit Ratio and NDCG. Ablations drop most of the gain when either the temporal embeddings or the counterfactual tuning component is removed.

What carries the argument

The central object is the temporal embedding $p_k$: a position vector added to every token of the $k$-th item, on top of ordinary token-level position embeddings, so each item becomes a cohesive temporal unit and later items carry stronger order cues. The twin-network counterfactual construction forms an order-erased input by replacing every item's $p_k$ with $p_0$, and the counterfactual tuning loss $\mathcal{L}_{CT} = \sum \ell(f_{\Theta+\Phi}(x_t,y_{<k}) - f_{\Theta+\Phi}(x_0,y_{<k}); y_k)$ is combined with the temporal-aware loss $\mathcal{L}_{TA}$ as $\mathcal{L} = \mathcal{L}_{TA} + \lambda \mathcal{L}_{CT}$. Sinusoidal and rotary position embeddings serve as two instantiations of $p_k$, with RoPE giving the stronger empirical results.

What would settle it

Run the released training code and inspect the actual form of $\mathcal{L}_{CT}$: if it does not match Eq. (8) because $f_{\Theta+\Phi}(x_t,y_{<k}) - f_{\Theta+\Phi}(x_0,y_{<k})$ is not a probability distribution, then the stated mechanism is not what produces the reported results. Alternatively, replace the difference with a proper KL divergence between the factual and counterfactual next-token distributions and re-run the three datasets; if accuracy is unchanged, the subtraction itself is not the active ingredient.

Watch

Extended reading notes

Core claim

The central claim is that temporal order should be treated as a causal factor separate from item content, and that doing so improves LLM-based sequential recommendation. CETRec injects order by adding the same position embedding to every token of each item in the history, and it trains the model with a counterfactual loss built from a twin-network model in which all items are placed at the first position, erasing relative and absolute order while keeping content fixed. Under this training, the paper reports that the RoPE variant of CETRec exceeds every baseline on all metrics on MovieLens, Steam, and LastFM, and that the counterfactual loss is responsible for a meaningful share of the gain.

Load-bearing premise

The load-bearing premise is that the counterfactual tuning loss subtracting two probability distributions is a well-defined training objective, since the paper does not say how to evaluate a loss on that difference; if the subtraction is not meaningful, the gains credited to counterfactual tuning are unsupported.

Editorial extensions

If this is right

  • On the three datasets tested, the full CETRec (RoPE) ranks above every baseline on HR@5, NDCG@5, HR@10, and NDCG@10, so the paper claims a clean win for order-aware LLM recommenders.
  • Dropping counterfactual tuning (w.o. CT) lowers performance across all datasets, so the counterfactual objective contributes beyond the temporal embeddings themselves.
  • Dropping both components leaves a model equivalent to the BIGRec baseline, so the reported gain over BIGRec is attributed to the two new components.
  • Reversing test sequences changes CETRec's scores more than it changes BIGRec's or CFT's scores, which the paper reads as evidence of higher temporal sensitivity.
  • The SinPE variant shows larger sensitivity to reversed order but lower accuracy than RoPE, implying temporal sensitivity and recommendation quality are not the same thing.

Reading between the lines

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

  • An editorial extension: the 'counterfactual erased' input is a single altered forward pass rather than a sampled counterfactual world, so the method functions less as causal inference and more as a regularizer that penalizes order-blind predictions; a natural check is whether simple order shuffling or temporal dropout produces the same gains.
  • Another extension: Eq. (8) subtracts two next-token probability vectors before applying the loss, which is not a valid probability distribution; re-implementing the objective with a proper divergence such as KL between factual and counterfactual predictions would reveal whether the subtraction itself matters.
  • A third extension: the same $p_k$ encodes both recency and ordinal distance, so the reported results do not separate absolute time from relative order; using time-interval or timestamp-difference embeddings would test which signal drives the improvement.
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. The paper proposes CETRec, a framework for improving temporal sensitivity in LLM-based sequential recommendation. It augments every token of each item with an item-level temporal embedding, constructs a counterfactual input by replacing all such embeddings with that of the first item, and trains with two losses: a temporal-aware cross-entropy loss L_TA on the ordered input, and a counterfactual tuning loss L_CT defined in Eq. (8) as a token-prediction loss on the difference between factual and counterfactual model outputs. Experiments on MovieLens, Steam, and LastFM compare CETRec against traditional and LLM-based baselines, reporting consistent improvements for the RoPE variant, ablation results showing the contribution of counterfactual tuning and temporal embeddings, and reversal experiments intended to measure temporal sensitivity.

Significance. If the counterfactual tuning objective is well-defined and behaves as claimed, CETRec would be a simple and general recipe for injecting temporal awareness into instruction-tuned LLM recommenders, with a released code base and evaluation on three public datasets. The consistent gains in Table 3 and the ablations in Table 4 are encouraging, and the paper explicitly frames falsifiable claims about absolute and relative order sensitivity that are partially tested through sequence reversal. However, the central training objective in Eq. (8) is underspecified as written, so the contribution of the paper's main novelty cannot currently be verified from the manuscript alone.

major comments (3)
  1. [§4.4, Eq. (8)] The counterfactual tuning loss is not well-defined. Section 4.4 states that f represents "the model's probability distribution over the next token," so f(x_t, y_<k) - f(x_0, y_<k) is a signed vector whose entries are not probabilities and whose components sum to zero. The loss function ell is never defined for such an input: the cross-entropy used in Eq. (1) requires a normalized probability vector, and no alternative such as a logit difference, a Bregman divergence, or a KL term is specified. This is load-bearing because L_CT is the only component that distinguishes CETRec from the w.o. CT ablation in Table 4, and the headline claim in §5.2 attributes the gains to this mechanism. The authors should either define ell precisely and state whether the argument to ell is probabilities, logits, or some other representation, or replace Eq. (8) with the actual objective used in the released code.
  2. [§4.3–§4.4, Eq. (7)] The causal inference section is presented as if it quantifies a causal effect, but no estimator for Eq. (7) is ever constructed. The twin network in Figure 3(b) is described graphically, yet the paper does not compute posterior distributions over the exogenous variables U, does not specify how P(Y*|T=t*, I=i, U=u) is obtained from the model, and does not connect the causal estimand to the training loss L_CT. As written, the causal analysis is motivational: it motivates why temporal order should matter, but it does not "precisely quantify" any causal effect. Please either state explicitly that Eq. (7) is a conceptual quantity that motivates the objective, or provide the actual identification and estimation steps.
  3. [§5.2, Table 3] All results are reported without standard deviations, number of seeds, or significance tests. Some of the claimed wins over the strongest baseline are modest in absolute terms, for example MovieLens NDCG@5 (0.0779 vs. LLaRA's 0.0710) and LastFM HR@10 (0.1030 vs. LLaRA's 0.0880). Given the claim in §5.2 that CETRec "outperforms all the baseline models on three datasets," the authors should provide variance information across multiple runs (or at least state that metrics are deterministic under a fixed seed) so that the reader can assess whether the differences are likely to be stable.
minor comments (5)
  1. [Throughout] The manuscript contains several typos and infelicities, including "real real-wrold" (§5.1.1), "summerized" (§5.1.1), "achiving" (§5.2), "Case Sduty" (§5.5), "performances better" (§5.2), and "mutipule baselines" (§6). These should be corrected.
  2. [§5.1.1] After describing MovieLens, Steam, and LastFM, the text says "For the two datasets, we construct interaction sequences of length 3 to 10," which is ambiguous because three datasets were just listed. Please clarify which two datasets this sentence refers to, or change it to "for the three datasets."
  3. [§5.4, Figure 4] The y-axis label "Change Rate" is not defined in the text. Please give the formula used to compute the performance change rate, for example (metric_reversed - metric_original) / metric_original, so that the reader can interpret the figure.
  4. [§4.4, Eq. (8)] The notation f_{\Theta+\Phi} is introduced in Eq. (2) with the explanation of LoRA, but Eq. (8) uses f_{\Theta+\Phi}(x_t, y_{<k}) without restating which parameters are active; this is understandable from context, but the paper would be clearer if the parameter set were defined once and reused consistently.
  5. [§5.2] The sentence "CETRec with RoPE outperforms all the baseline models on three datasets" should be qualified as applying to the chosen evaluation protocol and the three datasets used here, since the all-ranking protocol and the generation-to-item matching procedure can affect the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: CETRec's empirical gains are benchmarked externally; the counterfactual loss is a constructed objective, not a renamed input.

full rationale

The paper's derivation is empirical and self-contained. CETRec has two concrete training components: item-level temporal embeddings (Eq. 6) and the counterfactual tuning loss (Eq. 8), and its central claim is an empirical comparison on MovieLens, Steam, and LastFM against a range of baselines including both traditional and LLM-based methods. No fitted parameter is later renamed as a prediction: the temporal embedding p_k is assigned by position, the LoRA parameters are trained with L_TA + lambda*L_CT, and the reported HR/NDCG metrics are computed on held-out next-item prediction. The counterfactual loss is a constructed training objective, not a derived prediction of the evaluation metric; Section 5.4 uses held-out reversed sequences as an external sensitivity probe, so the 'temporal sensitivity' claim is not established by definition. The only basis for a self-citation concern is that two baselines (BIGRec and LLaRA) share authors with the present paper; they are used as comparison points rather than as load-bearing justification of CETRec's design, and several outside baselines (SASRec, GRU4Rec, Caser, P5, E4SRec, CFT) provide independent comparison. No uniqueness theorem or prior result by the same authors is invoked to rule out alternatives. The underspecification of ell in Eq. 8, where f is described as a probability distribution and then a difference of two such distributions is fed into a loss, is a correctness and reproducibility concern, not a circularity: nothing in the derivation reduces to an input by construction.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central empirical claim rests on the ad hoc construction of a counterfactual input and a tuned hyperparameter λ. The causal graph and embedding-erasure intervention are asserted rather than derived; no new physical or conceptual entities are introduced beyond the counterfactual input construction.

free parameters (1)
  • λ (counterfactual loss weight) = Tuned over {0.01, 0.1, 0.5, 1, 10}; final per-dataset value not reported
    Controls the trade-off between the temporal-aware loss and the counterfactual tuning loss; the reported improvements depend on this choice, and no sensitivity analysis is given.
assumptions (3)
  • ad hoc to paper Temporal order can be treated as an independent causal factor separable from item content (Figure 3a).
    The causal graph is asserted without empirical validation; the paper provides no evidence that T and I are separable causes of H.
  • ad hoc to paper Replacing all temporal embeddings with p_0 erases both absolute and relative temporal information while preserving all other factors.
    Section 4.3 defines the counterfactual intervention, but the paper does not verify that the embedding replacement fully removes temporal signals or that it does not corrupt item semantics.
  • ad hoc to paper The loss ℓ in Eq. (8) can accept a difference of model outputs as a valid training target.
    If f returns probabilities, the subtraction does not yield a probability distribution; the paper never defines ℓ for this case, leaving the training objective ambiguous.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Temporal Sensitivity of Large Language Model for Recommendation with Counterfactual Tuning." pith.science (2026). https://pith.science/paper/UHGYJBG4

@misc{pith2026250703047,
  author       = {Pith},
  title        = {Pith review of: Enhancing Temporal Sensitivity of Large Language Model for Recommendation with Counterfactual Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UHGYJBG4}},
  note         = {Machine review of arXiv:2507.03047}
}
read the original abstract

Recent advances have applied large language models (LLMs) to sequential recommendation, leveraging their pre-training knowledge and reasoning capabilities to provide more personalized user experiences. However, existing LLM-based methods fail to sufficiently leverage the rich temporal information inherent in users' historical interaction sequences, stemming from fundamental architectural constraints: LLMs process information through self-attention mechanisms that lack inherent sequence ordering and rely on position embeddings designed primarily for natural language rather than user interaction sequences. This limitation significantly impairs their ability to capture the evolution of user preferences over time and predict future interests accurately. To address this critical gap, we propose \underline{C}ounterfactual \underline{E}nhanced \underline{T}emporal Framework for LLM-Based \underline{Rec}ommendation (CETRec). CETRec is grounded in causal inference principles, which allow it to isolate and measure the specific impact of temporal information on recommendation outcomes. Combined with our counterfactual tuning task derived from causal analysis, CETRec effectively enhances LLMs' awareness of both absolute order (how recently items were interacted with) and relative order (the sequential relationships between items). Extensive experiments on real-world datasets demonstrate the effectiveness of our CETRec. Our code is available at https://anonymous.4open.science/r/CETRec-B9CE/.

Figures

Figures reproduced from arXiv: 2507.03047 by the authors.

Figure 1
Figure 1. Illustration of absolute vs. relative order effects. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The framework of CETRec. The user interaction sequence is transformed to the instruction tuning format. CETRec [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. (a) Causal graph for sequential recommendation [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Case study on Steam Dataset. signals, the latter contribute only so much—so the change rate cannot grow without bound. • CFT’s change rate is larger than BIGRec’s but still lower than CE￾TRec’s. This is because CFT improves the model’s understanding of the overall inte…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

67 extracted references · 45 canonical work pages

  1. [1]

    Qingyao Ai, Ting Bai, Zhao Cao, Yi Chang, Jiawei Chen, Zhumin Chen, Zhiyong Cheng, Shoubin Dong, Zhicheng Dou, Fuli Feng, Shen Gao, Jiafeng Guo, Xiangnan He, Yanyan Lan, Chenliang Li, Yiqun Liu, Ziyu Lyu, Weizhi Ma, Jun Ma, Zhaochun Ren, Pengjie Ren, Zhiqiang Wang, Mingwen Wang, Ji-Rong Wen, Le Wu, Xin Xin, Jun Xu, Dawei Yin, Peng Zhang, Fan Zhang, Weinan...

    work page 2023

  2. [2]

    Hervé-Madelein Attolou, Katerina Tzompanaki, Kostas Stefanidis, and Dimitris Kotzinos. 2024. Why-Not Explainable Graph Recommender. In ICDE. IEEE, 2245– 2257

    work page 2024

  3. [3]

    Alexander Balke and Judea Pearl. 1994. Probabilistic Evaluation of Counterfactual Queries. In AAAI. AAAI Press / The MIT Press, 230–237

    work page 1994

  4. [4]

    Keqin Bao, Jizhi Zhang, Wenjie Wang, Yang Zhang, Zhengyi Yang, Yancheng Luo, Fuli Feng, Xiangnan He, and Qi Tian. 2023. A Bi-Step Grounding Paradigm for Large Language Models in Recommendation Systems. CoRR abs/2308.08434 (2023)

    arXiv 2023

  5. [5]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He

  6. [6]

    Alex Beutel, Paul Covington, Sagar Jain, Can Xu, Jia Li, Vince Gatto, and Ed H. Chi. 2018. Latent Cross: Making Use of Context in Recurrent Recommender Systems. In WSDM. ACM, 46–54

    work page 2018

  7. [7]

    Iván Cantador, Peter Brusilovsky, and Tsvi Kuflik. 2011. Second workshop on information heterogeneity and fusion in recommender systems (HetRec2011). In RecSys. ACM, 387–388

    work page 2011

  8. [8]

    Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebas- tian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bra...

    work page 2023

Show all 67 references
  1. [9]

    Chi, and Minmin Chen

    Konstantina Christakopoulou, Madeleine Traverse, Trevor Potter, Emma Marriott, Daniel Li, Chris Haulk, Ed H. Chi, and Minmin Chen. 2020. Deconfounding User Satisfaction Estimation from Response Rate Bias. In RecSys. ACM, 450–455

    2020

  2. [10]

    Zhendong Chu, Zichao Wang, Ruiyi Zhang, Yangfeng Ji, Hongning Wang, and Tong Sun. 2024. Improve Temporal Awareness of LLMs for Sequential Recom- mendation. CoRR abs/2405.02778 (2024)

    2024 arXiv

  3. [11]

    Taylor, and Bill Ramsey

    Nick Craswell, Onno Zoeter, Michael J. Taylor, and Bill Ramsey. 2008. An experi- mental comparison of click position-bias models. In WSDM. ACM, 87–94

    2008

  4. [12]

    Qiang Cui, Shu Wu, Qiang Liu, Wen Zhong, and Liang Wang. 2020. MV-RNN: A Multi-View Recurrent Neural Network for Sequential Recommendation. IEEE Trans. Knowl. Data Eng. 32, 2 (2020), 317–331

    2020

  5. [13]

    Sunhao Dai, Ninglu Shao, Haiyuan Zhao, Weijie Yu, Zihua Si, Chen Xu, Zhongxi- ang Sun, Xiao Zhang, and Jun Xu. 2023. Uncovering ChatGPT’s Capabilities in Recommender Systems. In RecSys. ACM, 1126–1132

    2023

  6. [14]

    Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdh- ery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussai...

    2023

  7. [15]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ah- mad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sra- vankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston...

    2024 arXiv

  8. [16]

    Philipp Dufter, Martin Schmitt, and Hinrich Schütze. 2022. Position Information in Transformers: An Overview. Comput. Linguistics 48, 3 (2022), 733–763

    2022

  9. [17]

    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). arXiv preprint arXiv:2307.02046 (2023)

    2023 arXiv

  10. [18]

    Fenton, Martin Neil, and Anthony C

    Norman E. Fenton, Martin Neil, and Anthony C. Constantinou. 2020. The Book of Why: The New Science of Cause and Effect, Judea Pearl, Dana Mackenzie. Basic Books (2018). Artif. Intell. 284 (2020), 103286

    2018

  11. [19]

    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). In RecSys. ACM, 299–315

    2022

  12. [20]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

  13. [21]

    McAuley, and Wayne Xin Zhao

    Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian J. McAuley, and Wayne Xin Zhao. 2024. Large Language Models are Zero-Shot Rankers for Recommender Systems. In ECIR (2) (Lecture Notes in Computer Science, Vol. 14609). Springer, 364–381

    2024

  14. [22]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. In ICLR. OpenReview.net

    2022

  15. [23]

    Wenyue Hua, Shuyuan Xu, Yingqiang Ge, and Yongfeng Zhang. 2023. How to Index Item IDs for Recommendation Foundation Models. SIGIR-AP (2023)

    2023

  16. [24]

    Wen Huang, Lu Zhang, and Xintao Wu. 2022. Achieving Counterfactual Fairness for Causal Bandit. In AAAI. AAAI Press, 6952–6959

    2022

  17. [25]

    Matthew Jin, Syed Shahriar, Michele Tufano, Xin Shi, Shuai Lu, Neel Sundaresan, and Alexey Svyatkovskiy. 2023. InferFix: End-to-End Program Repair with LLMs. In ESEC/SIGSOFT FSE. ACM, 1646–1656

    2023

  18. [26]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In ICDM. 197–206

    2018

  19. [27]

    Sein Kim, Hongseok Kang, Kibum Kim, Jiwan Kim, Donghyun Kim, Minchul Yang, Kwangjin Oh, Julian McAuley, and Chanyoung Park. 2025. Lost in Sequence: Do Large Language Models Understand Sequential Recommendation? CoRR abs/2502.13909 (2025)

    2025 arXiv

  20. [28]

    Xinhang Li, Chong Chen, Xiangyu Zhao, Yong Zhang, and Chunxiao Xing. 2023. E4SRec: An Elegant Effective Efficient Extensible Solution of Large Language Models for Sequential Recommendation. CoRR abs/2312.02443 (2023)

    2023 arXiv

  21. [29]

    Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, and Xiang Wang. 2024. LLaRA: LLaRA: Large Language-Recommendation Assistant

    2024

  22. [31]

    Xinyu Lin, Wenjie Wang, Yongqi Li, Fuli Feng, See-Kiong Ng, and Tat-Seng Chua

  23. [32]

    Chang Liu, Chen Gao, Yuan Yuan, Chen Bai, Lingrui Luo, Xiaoyi Du, Xinlei Shi, Hengliang Luo, Depeng Jin, and Yong Li. 2022. Modeling Persuasion Factor of User Decision for Recommendation. In KDD. ACM, 3366–3376

    2022

  24. [33]

    CoRR abs/2310.06491 (2023)

    A Multi-facet Paradigm to Bridge Large Language Model and Recommen- dation. CoRR abs/2310.06491 (2023)

    2023 arXiv

  25. [34]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leik...

    2022

  26. [35]

    Judea Pearl. 2001. Direct and Indirect Effects. InUAI. Morgan Kaufmann, 411–420

    2001

  27. [36]

    Judea Pearl. 2009. Causality. Cambridge university press

    2009

  28. [37]

    Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. In ICLR (Poster). OpenReview.net

    2019

  29. [38]

    Steffen Rendle, Christoph Freudenthaler, and Lars Schmidt-Thieme. 2010. Factor- izing personalized Markov chains for next-basket recommendation. In WWW. ACM, 811–820

    2010

  30. [39]

    Jianlin Su, Murtadha H. M. Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. RoFormer: Enhanced transformer with Rotary Position Embedding. Neurocomputing 568 (2024), 127063

    2024

  31. [40]

    Juntao Tan, Shuyuan Xu, Wenyue Hua, Yingqiang Ge, Zelong Li, and Yongfeng Zhang. 2024. IDGenRec: LLM-RecSys Alignment with Textual ID Learning. In SIGIR. ACM, 355–364

    2024

  32. [41]

    Tran, Jonah Samost, Maciej Kula, Ed H

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Mahesh Sathiamoorthy. 2023. Recommender Systems with Generative Retrieval. In NeurIPS

    2023

  33. [42]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. In NIPS. 5998–6008

    2017

  34. [43]

    Mittelstadt, and Chris Russell

    Sandra Wachter, Brent D. Mittelstadt, and Chris Russell. 2017. Counterfactual Explanations without Opening the Black Box: Automated Decisions and the GDPR. CoRR abs/1711.00399 (2017)

    2017 arXiv

  35. [44]

    Wenjie Wang, Honghui Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, See- Kiong Ng, and Tat-Seng Chua. 2024. Learnable Item Tokenization for Generative Recommendation. In CIKM. ACM, 2400–2409

    2024

  36. [45]

    Jiaxi Tang and Ke Wang. 2018. Personalized top-n sequential recommendation via convolutional sequence embedding. In WSDM. 565–573

    2018

  37. [46]

    Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus

    Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. 2022. Emergent Abilities of Large Language Mod...

    2022

  38. [47]

    Wei Wei, Xubin Ren, Jiabin Tang, Qinyong Wang, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. 2024. LLMRec: Large Language Models with Graph Augmentation for Recommendation. In WSDM. ACM, 806–815

    2024

  39. [48]

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, et al . 2023. A Survey on Large Language Models for Recommendation. arXiv preprint arXiv:2305.19860 (2023)

    2023 arXiv

  40. [49]

    Wenjie Wang, Fuli Feng, Xiangnan He, Hanwang Zhang, and Tat-Seng Chua

  41. [50]

    Shuyuan Xu, Wenyue Hua, and Yongfeng Zhang. 2024. OpenP5: An Open-Source Platform for Developing, Training, and Evaluating LLM-based Recommender Systems. In SIGIR. ACM, 386–394

    2024

  42. [51]

    An Yan, Shuo Cheng, Wang-Cheng Kang, Mengting Wan, and Julian J. McAuley

  43. [52]

    Zhengyi Yang, Jiancan Wu, Yanchen Luo, Jizhi Zhang, Yancheng Yuan, An Zhang, Xiang Wang, and Xiangnan He. 2023. Large Language Model Can Interpret Latent Space of Sequential Recommender. CoRR abs/2310.20487 (2023)

    2023 arXiv

  44. [53]

    Jose, and Xiangnan He

    Fajie Yuan, Alexandros Karatzoglou, Ioannis Arapakis, Joemon M. Jose, and Xiangnan He. 2019. A Simple Convolutional Generative Network for Next Item Recommendation. In WSDM. ACM, 582–590

    2019

  45. [54]

    Kun Xiong, Wenwen Ye, Xu Chen, Yongfeng Zhang, Wayne Xin Zhao, Binbin Hu, Zhiqiang Zhang, and Jun Zhou. 2021. Counterfactual Review-based Recom- mendation. In CIKM. ACM, 2231–2240

    2021

  46. [55]

    Shengyu Zhang, Dong Yao, Zhou Zhao, Tat-Seng Chua, and Fei Wu. 2021. CauseRec: Counterfactual User Sequence Synthesis for Sequential Recommenda- tion. In SIGIR. ACM, 367–377

    2021

  47. [56]

    Sheng, Jiajie Xu, De- qing Wang, Guanfeng Liu, and Xiaofang Zhou

    Tingting Zhang, Pengpeng Zhao, Yanchi Liu, Victor S. Sheng, Jiajie Xu, De- qing Wang, Guanfeng Liu, and Xiaofang Zhou. 2019. Feature-level Deeper Self-Attention Network for Sequential Recommendation. InIJCAI. ijcai.org, 4320– 4326

    2019

  48. [57]

    Yang Zhang, Juntao You, Yimeng Bai, Jizhi Zhang, Keqin Bao, Wenjie Wang, and Tat-Seng Chua. 2024. Causality-Enhanced Behavior Sequence Modeling in LLMs for Personalized Recommendation. CoRR abs/2410.22809 (2024)

    2024 arXiv

  49. [58]

    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, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong W...

    2023 arXiv

  50. [59]

    Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation. In ICDE. IEEE, 1435–1448

    2024

  51. [60]

    Peiyan Zhang, Yuchen Yan, Xi Zhang, Liying Kang, Chaozhuo Li, Feiran Huang, Senzhang Wang, and Sunghun Kim. 2024. GPT4Rec: Graph Prompt Tuning for Streaming Recommendation. In SIGIR. ACM, 1774–1784

    2024

  52. [61]

    Yaochen Zhu, Jing Ma, Liang Wu, Qi Guo, Liangjie Hong, and Jundong Li. 2023. Path-Specific Counterfactual Fairness for Recommender Systems. In KDD. ACM, 3638–3649

    2023

  53. [62]

    Ziwei Zhu, Yun He, Yin Zhang, and James Caverlee. 2020. Unbiased Implicit Recommendation and Propensity Estimation via Combinational Joint Learning. In RecSys. ACM, 551–556

    2020

  54. [66]

    Haolin Zhou, Junwei Pan, Xinyi Zhou, Xihua Chen, Jie Jiang, Xiaofeng Gao, and Guihai Chen. 2024. Temporal Interest Network for User Response Prediction. In WWW (Companion Volume). ACM, 413–422

    2024

  55. [2016]

    Session-based Recommendations with Recurrent Neural Networks. In ICLR

  56. [2019]

    CosRec: 2D Convolutional Neural Networks for Sequential Recommenda- tion. In CIKM. ACM, 2173–2176

  57. [2021]

    In SIGIR

    Clicks can be Cheating: Counterfactual Recommendation for Mitigating Clickbait Issue. In SIGIR. ACM, 1288–1297

  58. [2023]

    In RecSys

    TALLRec: An Effective and Efficient Tuning Framework to Align Large Language Model with Recommendation. In RecSys. ACM, 1007–1014

  59. [2024]

    Bridging Items and Language: A Transition Paradigm for Large Language Model-Based Recommendation. In KDD. ACM, 1816–1826

Pith tools

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