Pith. sign in

REVIEW 4 major objections 4 minor 54 references

Improving Recommendation Fairness without Sensitive Attributes Using Multi-Persona LLMs

T0 review · 4 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 A recommender can be made fairer without sensitive attributes by using multi-persona LLMs to infer them, then training to discard that signal.

desk verdict A practically motivated and genuinely novel combination of multi-persona LLM inference and confusion-aware MI for fairness without demographics, but the main result is undercut by an unspecified extra loss term that appears only in the implementation details. read the letter →

arxiv 2505.19473 v1 pith:7U2RPTY7 submitted 2025-05-26 cs.IR

Haoran Xin , Ying Sun , Chao Wang , Yanke Yu , Weijia Zhang , Hui Xiong This is my paper · ORCID
classification cs.IR
keywords recommendationfairnesswithoutdemographicslargelanguagemodelsmulti-personainferencemutualinformationsensitiveattributeconfusionmatrixadversarialattackAUC
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

This paper tries to establish that a recommender can be made fairer even when no sensitive attributes are available at training time, by using large language models to infer those attributes from user behavior. The proposed framework, LLMFOSA, has multiple LLM annotators with different personas guess each user's sensitive attribute, a meta summarizer that produces rationales, and a confusion-aware module that turns these noisy inferences into robust sensitive-aware embeddings. The model is then optimized with mutual-information objectives so that recommendations use a sensitive-blind embedding that carries little to no sensitive information. If the claim is right, platforms that collect no demographic data can still reduce unfair treatment, a practical obstacle most fairness methods do not address.

What carries the argument

The load-bearing object is the pair of disentangled embeddings: a sensitive-aware embedding $s_u$ trained to retain the inferred sensitive information, and a sensitive-blind embedding $p_u$ used for ranking, optimized under the mutual-information objective $-I(S;A) + I(S;P) - I(P;R|S)$. The inference side is a three-agent pipeline: a persona editor generates diverse annotator personas, each annotator infers an attribute from the user's history, and a meta summarizer distills the rationales. The noise-handling side is a per-annotator learnable confusion matrix $F_i$ modeling $q(a_u^i = k | a_u = j)$, aligned across similar personas by a consensus term, plus a contrastive loss tying the sensitive embedding to the rationale embedding. The machinery's job is to make the inferred sensitive distribution reliable enough that optimizing $I(S;A)$ teaches the model what to remove rather than what to keep.

What would settle it

On a dataset where the LLM's inferred labels are known to be no better than chance, LLMFOSA's attacker AUC should not drop below the no-fairness baseline; if it does drop, the fairness improvement is an artifact of training dynamics rather than of the inferred sensitive signal. A direct check is to train the attack classifier on held-out true sensitive labels after LLMFOSA training: the paper's central claim predicts the attacker AUC stays near chance even though the model never saw true labels.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that noisy LLM-produced sensitive labels can substitute for true sensitive attributes in fairness training, and that the noise can be handled rather than avoided. With the matrix-factorization backbone on MovieLens-1M, the attacker's AUC for recovering gender drops from 0.7163 under a no-fairness model to 0.5974 for LLMFOSA, while N@20 stays near 0.23; on LastFM-360K the AUC drops from 0.6726 to 0.6305. The paper attributes this to multi-persona inference being substantially more accurate than clustering-based proxies, with LLM majority vote reaching 0.7644 accuracy on ML-1M versus 0.5818 for hierarchical clustering, and to the confusion-matrix correction preventing annotator errors from corrupting the learned sensitive representations.

Load-bearing premise

The method's value rests on LLM guesses about sensitive traits being close enough to the truth, and on those guesses' errors not being secretly linked to user behavior; if either fails, the model may still be unfair without anyone knowing.

Editorial extensions

If this is right

  • Platforms that collect no demographic data can still train fair recommenders: the fairness signal is purchased with LLM inference rather than with user disclosures.
  • The multi-persona plus confusion-correction design gives a general recipe for using LLM annotations as noisy labels, not only for gender or movies but for any sensitive attribute with a behavioral trace.
  • Because the sensitive-blind embedding is what ranking uses, the method yields a deployable model: after the offline LLM inference stage, the runtime model is the same lightweight collaborative filter.
  • The occupation-attribute experiments in the paper suggest the approach extends beyond binary attributes to multi-class sensitive categories.
  • The reported accuracy cost is modest, with N@20 within a few points of the no-fairness baseline, so the fairness gains are not bought by destroying recommendation quality.

Reading between the lines

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

  • The reliance of the LLM inference on cultural stereotypes cuts both ways: it may recover real attributes from behavior, but it may also encode stereotyped associations, so the fairness achieved is bounded by those associations and could fail on users whose behavior does not match them.
  • Since the confusion-matrix independence assumption is acknowledged in the paper as a limitation, a natural next step is instance-dependent confusion modeling; if annotator errors correlate with user behavior, the current correction could underestimate residual leakage.
  • A testable extension is to use the LLM's confidence or the agreement among personas as a per-user reliability weight rather than a single confusion matrix, which could improve fairness on harder-to-infer users.
  • The measured fairness is against a proxy of the sensitive attribute; an evaluation that trains the attacker on true labels for an unseen population would clarify whether the method removes true sensitive information or only the LLM's caricature of it.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes LLMFOSA, a framework for improving recommendation fairness when sensitive attributes are unavailable during training. It uses multiple LLM-powered annotators with distinct personas to infer sensitive attributes and rationales, builds robust sensitive-aware embeddings via confusion modeling and consensus regularization, and optimizes a mutual-information objective to separate sensitive-aware and sensitive-blind representations. Experiments on ML-1M and LFM-360K with MF and LightGCN backbones report large fairness gains measured by attacker AUC (e.g., from 0.7163 to 0.5974 on ML-1M with MF) at a modest accuracy cost, supported by ablations, parameter sensitivity studies, a different LLM backbone study, and an additional occupation-attribute evaluation.

Significance. If the result holds, the paper addresses a genuinely practical gap: improving recommendation fairness without access to demographic labels. The paper has several strengths: it evaluates fairness with held-out true sensitive attributes, provides ablations for the main components, includes a multi-LLM backbone study, and explicitly acknowledges limitations such as the independence assumption in Section I. However, the central empirical claim is currently undercut by an unspecified 'additional item-side MI lower bound' mentioned only in Section D, which is absent from the problem formulation, the derived losses, and the ablations. Without specifying and ablating that term, the reported gains cannot be attributed to the proposed method. The lack of variance reporting for the three-run averages in Table 1 further weakens the statistical support. These issues are fixable, so the work is worthy of major revision rather than rejection.

major comments (4)
  1. [Section D (Implementation Details)] The sentence 'An additional item-side MI lower bound was optimized due to superior practical performance' introduces an optimization term that appears nowhere in the problem formulation (Eq. 1), in the derived losses (Eqs. 2-14), or in any ablation. Since Table 1 is produced by a system whose training objective is strictly larger than the proposed and analyzed objective, the reader cannot determine whether the fairness gains come from the proposed sensitive-information modules or from this unspecified auxiliary term. Please define and derive the item-side bound, state explicitly in which training stage and loss it is used, and provide an ablation with and without it. This is load-bearing for the paper's central claim.
  2. [Table 1 and Section 4.2] The paper states that the reported numbers are means over three random runs, but no standard deviations, confidence intervals, or significance tests are provided. For example, the ML-1M/MF difference between LLMFOSA (AUC 0.5974) and NoFair (AUC 0.7163) is large, but several comparisons, such as LLMFOSA versus FairGNN on LFM-360K/GCN or versus AFSR on ML-1M/GCN, are smaller and could be within run-to-run variation. Please report variances or perform significance testing to support the claim that LLMFOSA 'outperforms all baselines in fairness.'
  3. [Section 3.2, Eqs. (4)-(5)] The derivation from the objective in Eq. (4) to the loss in Eq. (5) is not immediate. Eq. (4) contains the true attribute a_u, but the expectation is over (s_u, \tilde{a}_u^i), so a_u must be marginalized out. Eq. (5) is instead the log of the marginal probability p(\tilde{a}_u^i | s_u) under the confusion matrix and classifier, i.e., log \sum_j F^i_{j,\tilde{a}_u^i} C_j(s_u), which is not the same as the expectation of log q(\tilde{a}_u^i | a_u) q(a_u | s_u). Please provide the missing marginalization or revise the claimed bound so that the loss follows from the stated objective.
  4. [Appendix F and Eq. (12)] The derivation of the upper bound I(S;P) relies on an assumed Gaussian variational distribution q(s_u|p_u), and the lower bound for I(P;R|S) in Eq. (13) uses an arbitrary weighted cosine score function f(r_u, p_u, s_u) = \langle r_u, p_u + \alpha s_u \rangle. These choices are presented without justification or sensitivity analysis. Please state the role of these assumptions and provide at least a brief validation, since the MI bounds are the formal basis of the optimization objective.
minor comments (4)
  1. [Introduction, Section 4.5, Appendix G] There are several typos and grammatical issues, including 'Unfortunatly' in the Introduction, 'supurious' in Section 1, 'collborative' in Section 2.1, and 'Dicsussion' in the Appendix G heading. These should be corrected.
  2. [Section 4.5] The comparison between 'LLM-Single' and 'LLM-MV' is not exactly an ablation of the proposed multi-persona framework, since LLM-MV uses a majority vote rather than the confusion-aware learning described in Section 3.2. Please clarify the relationship between LLM-MV and the full LLMFOSA inference module.
  3. [Section D] The hyperparameter random search is described for K, N_o, \lambda_sim, and \lambda_fine, but the values of \lambda_ub and \lambda_lb in Eq. (14) are stated as fixed without explaining how they were chosen. Please report the search range or validation criterion for these weights as well.
  4. [Section 3.2] The notation for the inferred attribute set \tilde{A} and the true attribute set A is introduced informally; making the distinction explicit in the notation and in Eq. (2) would improve clarity, especially because the evaluation uses true A while training uses \tilde{A}.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the fairness objective is defined on LLM-inferred proxies and evaluated on held-out true attributes; no derivation step reduces to its own inputs.

full rationale

LLMFOSA's central objective (Eq. 1) uses the LLM-inferred sensitive attributes eA in place of the unavailable true A, and the fairness evaluation (Section C.2) trains an attacker on held-out true attributes; the reported AUC drop is therefore not forced by the training objective. The MI estimates are external, cited inequalities (CLUB [8] for the upper bound on I(S;P), conditional InfoNCE [14] for the lower bound on I(P;R|S)), and the confusion-matrix correction follows the standard Dawid-Skene assumption cited to [12,34], which the paper explicitly lists as a limitation in Section I. Equations (2)-(14) do not reuse the target fairness result as an input. The one substantive caveat is Section D: 'An additional item-side MI lower bound was optimized due to superior practical performance' is never specified or ablated, so the contribution of the stated objective versus the undisclosed term cannot be cleanly separated. This is a reproducibility/attribution gap, not a construction-level circularity, so the circularity score is 0.

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

The framework rests on a handful of tuned hyperparameters (N_o, K, λ_sim, λ_fine, etc.) and on the assumption that LLM-inferred labels faithfully represent the true sensitive attribute. The independence assumption for confusion matrices is explicitly called out as a limitation. No new physical or mathematical entities are postulated; the LLM agents, personas, and confusion matrices are computational constructs.

free parameters (6)
  • Number of annotators N_o = 4 (tuned over {4,6,8,10})
    Hyperparameter selected by random search on validation set; affects fairness and computational cost.
  • Nearest neighbors K = 1-3 (tuned)
    Used in consensus regularization Eq (6); chosen on validation set.
  • Regularization weights λ_sim, λ_fine = 10^-2 to 10^-4 (tuned)
    Coefficients in loss Eq (10); selected by random search.
  • MI weights λ_ub, λ_lb = 0.01, 0.1
    Weights for upper/lower bound losses in Eq (14); set directly, not tuned.
  • α in score function = 0.1
    Weight for sensitive embedding in conditional InfoNCE Eq (13); set directly.
  • Item-side MI lower bound = not specified
    Additional loss added 'due to superior practical performance' without derivation or specification (Section D).
assumptions (5)
  • domain assumption Annotator independence: each annotator's inference is independent of user behavior given the true sensitive attribute
    Invoked in Section 3.2 and Section I to decompose Eq (2) into Eq (4); the paper acknowledges this assumption may not hold and plans instance-dependent matrices.
  • ad hoc to paper Gaussian variational distribution for q(s_u|p_u)
    In Appendix F.2, to derive the CLUB upper bound, the variational distribution is assumed Gaussian without justification.
  • standard math Standard MI variational bounds (InfoNCE, CLUB) hold
    Used in Appendix F; these are known bounds, but their tightness depends on model capacity.
  • domain assumption LLM-inferred sensitive attributes are reliable proxies for true attributes
    The entire framework depends on the quality of multi-persona LLM inference; Table 2 shows accuracy ~0.72-0.76, which is better than clustering but still imperfect.
  • domain assumption Sensitive attributes are binary/multi and inferable from behavior
    The method assumes gender (and occupation) can be inferred from movie/play history; this may not hold for other attributes or domains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Recommendation Fairness without Sensitive Attributes Using Multi-Persona LLMs." pith.science (2026). https://pith.science/paper/7U2RPTY7

@misc{pith2026250519473,
  author       = {Pith},
  title        = {Pith review of: Improving Recommendation Fairness without Sensitive Attributes Using Multi-Persona LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7U2RPTY7}},
  note         = {Machine review of arXiv:2505.19473}
}
read the original abstract

Despite the success of recommender systems in alleviating information overload, fairness issues have raised concerns in recent years, potentially leading to unequal treatment for certain user groups. While efforts have been made to improve recommendation fairness, they often assume that users' sensitive attributes are available during model training. However, collecting sensitive information can be difficult, especially on platforms that involve no personal information disclosure. Therefore, we aim to improve recommendation fairness without any access to sensitive attributes. However, this is a non-trivial task because uncovering latent sensitive patterns from complicated user behaviors without explicit sensitive attributes can be difficult. Consequently, suboptimal estimates of sensitive distributions can hinder the fairness training process. To address these challenges, leveraging the remarkable reasoning abilities of Large Language Models (LLMs), we propose a novel LLM-enhanced framework for Fair recommendation withOut Sensitive Attributes (LLMFOSA). A Multi-Persona Sensitive Information Inference module employs LLMs with distinct personas that mimic diverse human perceptions to infer and distill sensitive information. Furthermore, a Confusion-Aware Sensitive Representation Learning module incorporates inference results and rationales to develop robust sensitive representations, considering the mislabeling confusion and collective consensus among agents. The model is then optimized by a formulated mutual information objective. Extensive experiments on two public datasets validate the effectiveness of LLMFOSA in improving fairness.

Figures

Figures reproduced from arXiv: 2505.19473 by the authors.

Figure 1
Figure 1. The overview of the LLMFOSA framework. 1) LLM-powered agents collaboratively [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Contributions of different compo￾nents in LLMFOSA to the fairness. (a) ML-1M (b) LFM-360K [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 4
Figure 4. Parameter sensitivity on two hyper-parameters. Results of accuracy are in Section E.1. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Contributions of different components in [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 7
Figure 7. Figure 7: The fairness of LLMFOSA when equipped with different LLM backbones. We also evaluate the performance of LLM￾FOSA using different LLM backbones. Specifi￾cally, we experiment with two widely used open￾source LLMs, i.e., Qwen2.5-7B6 and Gemma2- 2B7 , alongside Llama3-8B, …
Figure 8
Figure 8. Figure 8: An example of the meta summarizer’s response. [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: An example of the annotator’s response. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 42 canonical work pages

  1. [1]

    Learning optimal and fair decision trees for non-discriminative decision-making

    Sina Aghaei, Mohammad Javad Azizi, and Phebe Vayanos. Learning optimal and fair decision trees for non-discriminative decision-making. InProceedings of the AAAI conference on artificial intelligence, volume 33, pages 1418–1426, 2019

    work page 2019

  2. [2]

    Fairness without demographic data: A survey of approaches

    Carolyn Ashurst and Adrian Weller. Fairness without demographic data: A survey of approaches. InProceedings of the 3rd ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization, pages 1–12, 2023

    work page 2023

  3. [3]

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

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan 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, pages 1007–1014, 2023

    2023

  4. [4]

    Compositional fairness constraints for graph embeddings

    Avishek Bose and William Hamilton. Compositional fairness constraints for graph embeddings. InInternational Conference on Machine Learning, pages 715–724. PMLR, 2019

    work page 2019

  5. [5]

    Universitat Pompeu Fabra, 2009

    Òscar Celma Herrada et al.Music recommendation and discovery in the long tail. Universitat Pompeu Fabra, 2009

    work page 2009

  6. [6]

    Fairness without demographics through knowledge distillation.Advances in Neural Information Processing Systems, 35:19152–19164, 2022

    Junyi Chai, Taeuk Jang, and Xiaoqian Wang. Fairness without demographics through knowledge distillation.Advances in Neural Information Processing Systems, 35:19152–19164, 2022

    work page 2022

  7. [7]

    Improving recommendation fairness via data augmentation

    Lei Chen, Le Wu, Kun Zhang, Richang Hong, Defu Lian, Zhiqiang Zhang, Jun Zhou, and Meng Wang. Improving recommendation fairness via data augmentation. InProceedings of the ACM Web Conference 2023, pages 1012–1020, 2023

    work page 2023

  8. [8]

    Club: A contrastive log-ratio upper bound of mutual information

    Pengyu Cheng, Weituo Hao, Shuyang Dai, Jiachang Liu, Zhe Gan, and Lawrence Carin. Club: A contrastive log-ratio upper bound of mutual information. InInternational conference on machine learning, pages 1779–1788. PMLR, 2020

    work page 2020

Show all 54 references
  1. [9]

    Flexibly fair representation learning by disentanglement

    Elliot Creager, David Madras, Jörn-Henrik Jacobsen, Marissa Weis, Kevin Swersky, Toniann Pitassi, and Richard Zemel. Flexibly fair representation learning by disentanglement. In International conference on machine learning, pages 1436–1445. PMLR, 2019

    2019

  2. [10]

    Environment inference for invariant learning

    Elliot Creager, Jörn-Henrik Jacobsen, and Richard Zemel. Environment inference for invariant learning. InInternational Conference on Machine Learning, pages 2189–2200. PMLR, 2021

    2021

  3. [11]

    Say no to the discrimination: Learning fair graph neural networks with limited sensitive attribute information

    Enyan Dai and Suhang Wang. Say no to the discrimination: Learning fair graph neural networks with limited sensitive attribute information. InProceedings of the 14th ACM International Conference on Web Search and Data Mining, pages 680–688, 2021

    2021

  4. [12]

    Maximum likelihood estimation of observer error-rates using the em algorithm.Journal of the Royal Statistical Society: Series C (Applied Statistics), 28(1):20–28, 1979

    Alexander Philip Dawid and Allan M Skene. Maximum likelihood estimation of observer error-rates using the em algorithm.Journal of the Royal Statistical Society: Series C (Applied Statistics), 28(1):20–28, 1979

    1979

  5. [13]

    Fairness through awareness

    Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. Fairness through awareness. InProceedings of the 3rd innovations in theoretical computer science conference, pages 214–226, 2012

    2012

  6. [14]

    Controllable guarantees for fair outcomes via contrastive information estimation

    Umang Gupta, Aaron M Ferber, Bistra Dilkina, and Greg Ver Steeg. Controllable guarantees for fair outcomes via contrastive information estimation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 7610–7619, 2021

    2021

  7. [15]

    Equality of opportunity in supervised learning

    Moritz Hardt, Eric Price, and Nati Srebro. Equality of opportunity in supervised learning. Advances in neural information processing systems, 29, 2016

    2016

  8. [16]

    The movielens datasets: History and context.Acm transactions on interactive intelligent systems (tiis), 5(4):1–19, 2015

    F Maxwell Harper and Joseph A Konstan. The movielens datasets: History and context.Acm transactions on interactive intelligent systems (tiis), 5(4):1–19, 2015

    2015

  9. [17]

    Fairness without demographics in repeated loss minimization

    Tatsunori Hashimoto, Megha Srivastava, Hongseok Namkoong, and Percy Liang. Fairness without demographics in repeated loss minimization. InInternational Conference on Machine Learning, pages 1929–1938. PMLR, 2018. 10

    1929

  10. [18]

    Lightgcn: Simplifying and powering graph convolution network for recommendation

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. Lightgcn: Simplifying and powering graph convolution network for recommendation. InProceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages ...

    2020

  11. [19]

    Music personalization at spotify

    Kurt Jacobson, Vidhya Murali, Edward Newett, Brian Whitman, and Romain Yon. Music personalization at spotify. InProceedings of the 10th ACM Conference on Recommender Systems, pages 373–373, 2016

    2016

  12. [20]

    Ir evaluation methods for retrieving highly relevant documents

    Kalervo Järvelin and Jaana Kekäläinen. Ir evaluation methods for retrieving highly relevant documents. InACM SIGIR Forum, volume 51, pages 243–250. ACM New York, NY , USA, 2017

    2017

  13. [21]

    Fairness without demographics through adversarially reweighted learning

    Preethi Lahoti, Alex Beutel, Jilin Chen, Kang Lee, Flavien Prost, Nithum Thain, Xuezhi Wang, and Ed Chi. Fairness without demographics through adversarially reweighted learning. Advances in neural information processing systems, 33:728–740, 2020

    2020

  14. [22]

    Algorithmic bias? an empirical study of apparent gender-based discrimination in the display of stem career ads.Management science, 65(7): 2966–2981, 2019

    Anja Lambrecht and Catherine Tucker. Algorithmic bias? an empirical study of apparent gender-based discrimination in the display of stem career ads.Management science, 65(7): 2966–2981, 2019

    2019

  15. [23]

    Towards person- alized fairness based on causal notion

    Yunqi Li, Hanxiong Chen, Shuyuan Xu, Yingqiang Ge, and Yongfeng Zhang. Towards person- alized fairness based on causal notion. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1054–1063, 2021

    2021

  16. [24]

    Llara: Large language-recommendation assistant

    Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. Llara: Large language-recommendation assistant. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1785– 1795, 2024

    2024

  17. [25]

    Rella: Retrieval-enhanced large language models for lifelong sequential behavior comprehension in recommendation

    Jianghao Lin, Rong Shan, Chenxu Zhu, Kounianhua Du, Bo Chen, Shigang Quan, Ruiming Tang, Yong Yu, and Weinan Zhang. Rella: Retrieval-enhanced large language models for lifelong sequential behavior comprehension in recommendation. InProceedings of the ACM on Web Conference 2024...

    2024

  18. [26]

    Reasoning on graphs: Faithful and interpretable large language model reasoning.arXiv preprint arXiv:2310.01061, 2023

    Linhao Luo, Yuan-Fang Li, Gholamreza Haffari, and Shirui Pan. Reasoning on graphs: Faithful and interpretable large language model reasoning.arXiv preprint arXiv:2310.01061, 2023

    2023 arXiv

  19. [27]

    Visual classification via description from large language models.arXiv preprint arXiv:2210.07183, 2022

    Sachit Menon and Carl V ondrick. Visual classification via description from large language models.arXiv preprint arXiv:2210.07183, 2022

    2022 arXiv

  20. [28]

    Invariant representations without adversarial training.Advances in neural information processing systems, 31, 2018

    Daniel Moyer, Shuyang Gao, Rob Brekelmans, Aram Galstyan, and Greg Ver Steeg. Invariant representations without adversarial training.Advances in neural information processing systems, 31, 2018

    2018

  21. [29]

    Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018

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

    2018 arXiv

  22. [30]

    A theory of justice

    John Rawls. A theory of justice. InApplied ethics, pages 21–29. Routledge, 2017

    2017

  23. [31]

    Representation learning with large language models for recommendation

    Xubin Ren, Wei Wei, Lianghao Xia, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. Representation learning with large language models for recommendation. In Proceedings of the ACM on Web Conference 2024, pages 3464–3475, 2024

    2024

  24. [32]

    Bpr: Bayesian personalized ranking from implicit feedback.arXiv preprint arXiv:1205.2618, 2012

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme. Bpr: Bayesian personalized ranking from implicit feedback.arXiv preprint arXiv:1205.2618, 2012

    2012 arXiv

  25. [33]

    Leveraging large language models for multiple choice question answering

    Joshua Robinson and David Wingate. Leveraging large language models for multiple choice question answering. InThe Eleventh International Conference on Learning Representations, 2022

    2022

  26. [34]

    Deep learning from crowds

    Filipe Rodrigues and Francisco Pereira. Deep learning from crowds. InProceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 11

    2018

  27. [35]

    Opening the black box of deep neural networks via information.arXiv preprint arXiv:1703.00810, 2017

    Ravid Shwartz-Ziv and Naftali Tishby. Opening the black box of deep neural networks via information.arXiv preprint arXiv:1703.00810, 2017

    2017 arXiv

  28. [36]

    Learning controllable fair representations

    Jiaming Song, Pratyusha Kalluri, Aditya Grover, Shengjia Zhao, and Stefano Ermon. Learning controllable fair representations. InThe 22nd International Conference on Artificial Intelligence and Statistics, pages 2164–2173. PMLR, 2019

    2019

  29. [37]

    A survey on the fairness of recommender systems.ACM Transactions on Information Systems, 41(3):1–43, 2023

    Yifan Wang, Weizhi Ma, Min Zhang, Yiqun Liu, and Shaoping Ma. A survey on the fairness of recommender systems.ACM Transactions on Information Systems, 41(3):1–43, 2023

    2023

  30. [38]

    Can small language models be good reasoners for sequential recommendation? InProceedings of the ACM on Web Conference 2024, pages 3876–3887, 2024

    Yuling Wang, Changxin Tian, Binbin Hu, Yanhua Yu, Ziqi Liu, Zhiqiang Zhang, Jun Zhou, Liang Pang, and Xiao Wang. Can small language models be good reasoners for sequential recommendation? InProceedings of the ACM on Web Conference 2024, pages 3876–3887, 2024

    2024

  31. [39]

    Llmrec: Large language models with graph augmentation for recommendation

    Wei Wei, Xubin Ren, Jiabin Tang, Qinyong Wang, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin, and Chao Huang. Llmrec: Large language models with graph augmentation for recommendation. InProceedings of the 17th ACM International Conference on Web Search and Data Mining, pages 8...

    2024

  32. [40]

    Learning fair representations for recommendation: A graph-based perspective

    Le Wu, Lei Chen, Pengyang Shao, Richang Hong, Xiting Wang, and Meng Wang. Learning fair representations for recommendation: A graph-based perspective. InProceedings of the Web Conference 2021, pages 2198–2208, 2021

    2021

  33. [41]

    Fair class balancing: Enhancing model fairness without observing sensitive attributes

    Shen Yan, Hsien-te Kao, and Emilio Ferrara. Fair class balancing: Enhancing model fairness without observing sensitive attributes. InProceedings of the 29th ACM International Conference on Information & Knowledge Management, pages 1715–1724, 2020

    2020

  34. [42]

    Beyond parity: Fairness objectives for collaborative filtering

    Sirui Yao and Bert Huang. Beyond parity: Fairness objectives for collaborative filtering. Advances in neural information processing systems, 30, 2017

    2017

  35. [43]

    Fair sequential recommendation without user demographics

    Huimin Zeng, Zhankui He, Zhenrui Yue, Julian McAuley, and Dong Wang. Fair sequential recommendation without user demographics. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 395–404, 2024

    2024

  36. [44]

    Siren’s song in the ai ocean: a survey on hallucination in large language models.arXiv preprint arXiv:2309.01219, 2023

    Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al. Siren’s song in the ai ocean: a survey on hallucination in large language models.arXiv preprint arXiv:2309.01219, 2023

    2023 arXiv

  37. [45]

    Fairlisa: Fair user modeling with limited sensitive attributes information

    Zheng Zhang, Qi Liu, Hao Jiang, Fei Wang, Yan Zhuang, Le Wu, Weibo Gao, and Enhong Chen. Fairlisa: Fair user modeling with limited sensitive attributes information. InThirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  38. [46]

    Fair representation learning for recommendation: A mutual information perspective

    Chen Zhao, Le Wu, Pengyang Shao, Kun Zhang, Richang Hong, and Meng Wang. Fair representation learning for recommendation: A mutual information perspective. InProceedings of the AAAI Conference on artificial intelligence, volume 37, pages 4911–4919, 2023

    2023

  39. [47]

    Towards fair classifiers without sensitive attributes: Exploring biases in related features

    Tianxiang Zhao, Enyan Dai, Kai Shu, and Suhang Wang. Towards fair classifiers without sensitive attributes: Exploring biases in related features. InProceedings of the Fifteenth ACM International Conference on Web Search and Data Mining, pages 1433–1442, 2022

    2022

  40. [48]

    Adaptive fair representation learning for personalized fairness in recommendations via information alignment

    Xinyu Zhu, Lilin Zhang, and Ning Yang. Adaptive fair representation learning for personalized fairness in recommendations via information alignment. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 427–436, 2024

    2024

  41. [49]

    Collaborative large language model for recommender systems

    Yaochen Zhu, Liang Wu, Qi Guo, Liangjie Hong, and Jundong Li. Collaborative large language model for recommender systems. InProceedings of the ACM on Web Conference 2024, pages 3162–3172, 2024. 12 Appendix A Prompts Here we provide all the prompts used in the multi-persona sen...

    2024

  42. [50]

    # ! "" #

    FairMI[ 46] defines fairness with a two-fold mutual information objective and utilizes bounds for model optimization;recommendation fairness methods with limited or no sensitive attributes:1) FairGNN[ 11] introduces a sensitive estimator to predict missing sensitive attributes...

    2011

  43. [51]

    These films areoften associated with a female audienceand suggest a fondness for traditional fairy tales and romance

    The presence of classic Disney movies like Snow White, Cinderella, and Beauty and the Beast is a common thread throughout the annotations. These films areoften associated with a female audienceand suggest a fondness for traditional fairy tales and romance

  44. [52]

    The inclusion of romantic comedies and dramas, such as My Fair Lady, Meet Joe Black, and Titanic, is another consistent theme.These genres often appeal to womenand suggest an interest in relationships and emotional storytelling

  45. [53]

    4.The absence of action-oriented or sci-fi movies, which are often popular among male audiences, is a notable pattern

    The presence of family-friendly movies like Toy Story, James and the Giant Peach, and Mulan, as well as animated films like Aladdin, Hercules, and Bambi,suggests a love for lighthearted, feel-good stories and a willingness to engage with popular culture. 4.The absence of actio...

  46. [54]

    While it’s possible that a male user could have similar tastes, the consistency of these themes and patterns across the annotations suggests that the user is likely a female

    The overall tone of the movie list is characterized by a focus on romance, relationships, and family-friendly entertainment, which isoften associated with a female audience. While it’s possible that a male user could have similar tastes, the consistency of these themes and pat...

Pith tools

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