Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Bridging Jensen Gap for Max-Min Group Fairness Optimization in Recommendation

T0 review · 5 major / 6 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 This paper shows that mini-batch training of max-min fair recommender models suffers a Jensen gap, and introduces FairDual, a dual optimization algorithm that bounds the gap and converges sub-linearly to the global optimum of the…

desk verdict Solid empirical package, but the theoretical core—Theorem 1's reformulation and Theorem 4's rate—does not hold up; treat as an engineering paper, not a theory paper. read the letter →

arxiv 2502.09319 v1 pith:6NLWO2WF submitted 2025-02-13 cs.IR cs.LG

Chen Xu , Yuxin Li , Wenjie Wang , Liang Pang , Jun Xu , Tat-Seng Chua This is my paper · ORCID
classification cs.IRcs.LG
keywords groupmax-minfairnessrecommendersystemsJensengapmini-batchsamplingdualoptimizationmirrorgradientdescentfairness-accuracytrade-offshadowprice
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 optimizing group max-min fairness in recommender systems under mini-batch sampling breaks the usual assumption of sample independence, making the loss nonlinear and non-additive across batches. Because of that nonlinearity, a Jensen gap appears: the model's convergence point drifts away from the true optimum, and the gap worsens as batch size shrinks and the number of item groups grows. The paper further claims the constrained objective can be rewritten as a group-weighted accuracy objective, then gives FairDual, a dual mirror-gradient algorithm that learns per-group weights and, under random-shuffle mini-batch training, bounds the Jensen gap by $O(B^{-1/2})$ with batch size $B$. If these claims are right, FairDual offers a theoretically grounded, scalable way to keep both accuracy and worst-off-group fairness when training large recommender models on small batches. The paper supports the claims with experiments on MIND, Amazon-Book, and Amazon-Electronic across six recommendation backbones.

What carries the argument

The load-bearing construction is a dual reformulation: the max-min fairness constraint is turned into a group-weight vector $s_g = 1 - \mu_g$, where $\mu$ is a shadow price optimized by mirror gradient descent. The argument starts from Theorem 1, which asserts that the constrained objective equals $\min_w b^{\top}(\hat{A}^{\top} w)^{1+t}$ for some $t \geq 0$ and weight vector $b$, making the loss a nonlinear power of group utilities and thus subject to Jensen's inequality. FairDual then alternates between primal updates on the weighted loss $(s_j)^{\top} l_j$ and dual updates $\mu_j = \arg\min_{\mu \in M}(g_j^{\top} \mu + \eta \|\mu - \mu_{j-1}\|^2)$ over the convex feasible set $M = \{\mu : \sum_{g \in S} \mu_g m_g \geq -\lambda, \forall S \subseteq G\}$, with gradient estimates built from sampled item embeddings of the row-normalized item-group adjacency matrix $\hat{A}$. This machinery carries the proof that the complementary slackness error, and hence the Jensen gap, is bounded at a sub-linear rate.

What would settle it

Solve the original constrained objective in Equation (1) exactly on a small dataset with two item groups and a handful of users, then compare its minimum to the minimum of $\min_w b^{\top}(\hat{A}^{\top} w)^{1+t}$ over $t \in [0, \infty)$; if no $t$ matches, or if the empirical Jensen gap under FairDual does not shrink like $B^{-1/2}$ as batch size grows, the central theorem is falsified.

Watch

Extended reading notes

Core claim

The central claim is that the max-min fairness constrained recommendation objective, when trained with mini-batches, is not linearly additive, so the attained solution differs from the globally optimal solution by a Jensen gap $J(B) = |L_B - L|$ that grows with smaller batch sizes and larger group sizes. For the proposed FairDual algorithm, the paper proves a concrete bound: $J(B) \leq \frac{H}{\eta} + \frac{|U|L|G|^2}{B(1-\alpha)\sigma\eta} + \frac{L|G|^2}{2(1-\alpha)^2\sigma\eta}$, and with learning rate $\eta = O(B^{-1/2})$ this bound is comparable to $O(B^{-1/2})$. That sub-linear bound is the paper's headline result: under random-shuffle mini-batch training, FairDual converges to the globally optimal point of the MMF-constrained objective, and the optimization gap from batch sampling is kept under control even with small batches and many groups. Empirically, the paper reports that FairDual improves NDCG, MRR, and the MMF metric over reweighting and debiasing baselines across six backbones.

Load-bearing premise

Everything downstream depends on the assertion in Theorem 1 that some exponent $t \geq 0$ and weight vector $b$ exist making the MMF-constrained objective exactly equal to $\min_w b^{\top}(\hat{A}^{\top} w)^{1+t}$; the appendix argues this by continuity but never exhibits $t$ or proves it exists on the actual feasible set.

Editorial extensions

If this is right

  • If the Jensen gap bound is correct, FairDual converges at a sub-linear rate $O(B^{-1/2})$ to the globally optimal point of the MMF-constrained objective under random-shuffle mini-batch training, rather than to a biased fixed point.
  • With the gap bounded, small mini-batch sizes and large group counts no longer cause the fairness-accuracy trade-off to drift uncontrollably, which matters for industrial recommender systems that must train on tiny batches at scale.
  • The dual weights $s_g = 1 - \mu_g$ provide a principled, per-group reweighting scheme that can be plugged into existing backbone models without changing the underlying architecture or loss.
  • The reformulation extends to user-side and two-sided fairness by swapping the item-group adjacency matrix for the user-side equivalent and adding a second dual variable, so the same analysis transfers to more complex fairness settings.
  • Because the paper connects the objective to the power-mean welfare family, the same dual machinery could be applied to other nonlinear fairness functions, such as entropy-based or $\alpha$-fairness objectives.

Reading between the lines

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

  • The $O(B^{-1/2})$ rate suggests a testable design rule: if the empirical Jensen gap under FairDual does not shrink roughly in proportion to $1/\sqrt{B}$ as batch size grows, the theoretical picture is incomplete for real nonconvex recommender models.
  • Reading $\mu$ as a shadow price opens an extension the paper leaves implicit: the dual variable could be used to enforce time-varying fairness budgets, such as weekly exposure caps for provider groups, without retraining the primal model.
  • Theorem 1's existence claim for $t$ is nonconstructive, so an empirical check on synthetic two-group problems would clarify whether the reformulation holds outside the datasets tested in the paper; such a check is a natural extension of the paper's own simulation analysis.
  • The same Jensen gap diagnostic could be applied to other non-additive ranking objectives, such as Gini welfare or exposure concentration constraints, turning batch-size analysis into a general tool for fairness optimization in recommendation.
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

5 major / 6 minor

Summary. The paper studies group max-min fairness (MMF) in recommender systems and argues that adding an MMF constraint to the training objective breaks sample independence, causing a Jensen gap when mini-batch sampling is used. The authors claim that the MMF-constrained objective can be reformulated as a weighted power-sum objective, that this reformulation introduces a Jensen gap, and that their proposed algorithm FairDual—based on dual mirror-gradient descent—can bound this gap at a sub-linear rate. The empirical section compares FairDual against several fairness baselines on three datasets with multiple backbone models. The central theoretical contributions are Theorem 1 (equivalence to a weighted power-sum objective), Theorem 2 (existence and monotonicity of the Jensen gap), Theorem 3 (dual reformulation as group re-weighting), and Theorem 4 (a bound on the Jensen gap for FairDual).

Significance. If the theoretical claims were correct, the paper would make a meaningful contribution to fairness-aware recommendation by identifying a practical source of optimization bias (the Jensen gap) and providing a scalable algorithm with a theoretical guarantee. The experimental study is extensive, covering several recommender backbones and datasets, and the authors release their data and code, which is commendable for reproducibility. However, the significance is undermined by the fact that the main theoretical results are not established: Theorem 1 is asserted via an invalid continuity argument, Theorem 3 only shows weak duality, and Theorem 4's claimed rate is algebraically inconsistent. The conceptual message—that non-linear fairness objectives create a mini-batch bias—is reasonable, but the paper's specific formal claims and the FairDual guarantee rest on unproven or incorrect derivations.

major comments (5)
  1. [§4.1, Theorem 1 and Appendix A]
  2. [Appendix B, Lemma 1 and Appendix C, Theorem 2]
  3. [§5.1, Theorem 3 and Appendix F]
  4. [§5.2.3, Theorem 4, Equation (10)]
  5. [Appendix A, derivation from Equation (1) to Equations (11)–(12)]
minor comments (6)
  1. [Abstract and §7]
  2. [Table 1, Electronic K=20]
  3. [§4.1, Theorem 1 statement]
  4. [§4.1, Theorem 2 and Figure 1]
  5. [Eq. (3) and surrounding text]
  6. [Appendix I, dataset statistics]

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reduction found; the main theoretical risk is an unproven reformulation (Theorem 1) and a likely algebraic slip in Theorem 4, which are correctness concerns rather than self-referential reasoning.

full rationale

The derivation chain is: Eq. (1), the MMF-constrained loss, is said (Theorem 1) to be optimizable as L = min_w b^T(Â^T w)^(1+t) for existential t and b; Theorem 2 defines the mini-batch Jensen gap via the same power function f(x)=x^(1+t); Theorem 3 derives the dual re-weighting form; Theorem 4 bounds J(B) by a mirror-descent regret argument. The only step that could look circular is Theorem 1: Appendix A asserts the existence of t,b by a continuity/intermediate-value argument and admits 'the specific value of t is an implicit function and cannot be solved explicitly in closed form.' That is a serious missing proof, and the subsequent Jensen-gap analysis is conditional on an equivalence that is never instantiated. But it is not a circular reduction: Eq. (2) is not defined to be Eq. (1), no fitted parameter is relabeled as a prediction, and the proof does not rely on the paper's own prior work. Theorem 3's dual reformulation is a standard Lagrangian derivation supported by the external mirror-descent result of Balseiro et al. (2021). Self-citations (e.g., Xu et al. 2023 in Section 4.2; Xu et al. 2024b in Appendix I) are used for simulation parameters and dataset preprocessing, not to justify the convergence guarantee. Theorem 4 also appears to contain an algebraic error: substituting eta = O(B^{-1/2}) into Eq. (10) leaves terms of order O(B^{1/2}), not O(B^{-1/2}); this is a correctness defect, not circularity. I therefore find no specific step in which a claimed prediction reduces by construction to its inputs.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central argument rests on an unproven existential reformulation, an unproven strong-duality step, and bound constants that are never instantiated. The method also introduces several tunable hyperparameters (λ, α, η, m_g). This ledger makes explicit that the theoretical guarantee is conditional on assumptions that the paper does not discharge.

free parameters (7)
  • t (exponent in Theorem 1) = not computed, asserted to exist
    Introduced in Theorem 1 to make the max-min objective equal to b^T(Â^T w)^{1+t}. The entire Jensen gap analysis depends on this exponent; the paper admits it has no closed form.
  • b (weight vector in Theorem 1) = not specified
    A vector b ∈ R^{|G|} is asserted to exist alongside t; its construction is not given.
  • λ (fairness trade-off coefficient) = tuned in [0,10]
    Controls the trade-off between accuracy and MMF; selected on validation data.
  • m_g (group weight) = set to group size |I_g|
    A modeling choice for the per-group 'minimum wage'; could be chosen differently.
  • η (dual learning rate) = tuned in [1e-5,5e-3]
    Dual gradient step size; the convergence bound depends on it.
  • α (momentum coefficient) = not specified in main text
    Momentum weight in the dual gradient update; appears in the Jensen gap bound.
  • H, L, σ (bound constants in Theorem 4) = not instantiated
    The theorem assumes existence of H>0, L>0, and σ>0 but never gives values, so the bound is not numerically evaluable.
assumptions (4)
  • ad hoc to paper There exist t ∈ [0,∞) and b ∈ R^{|G|} such that Equation (1) equals min_w b^T(Â^T w)^{1+t} (Theorem 1).
    The proof in Appendix A uses continuity and an intermediate value argument but never shows the required t is attained; this is the foundation of the Jensen gap analysis.
  • domain assumption Strong duality holds for the Lagrangian in Theorem 3, i.e., the inequality in Eq. (20) is an equality.
    The derivation shows only max-min ≤ min-max; zero duality gap is not proven, so the reweighting reformulation may be an upper bound rather than an equivalent objective.
  • domain assumption Group utilities are bounded, e_g ≤ L, and per-user ranking scores are bounded by K, giving bounded subgradients in Theorem 4.
    Needed to justify ||eg_j||_2^2 ≤ L|G|^2 and the application of Balseiro et al.'s Theorem 2.
  • standard math The feasible region M for the dual variable is convex and the projection step is efficient.
    Lemma 3 proves convexity; projection relies on cvxpy. This is standard.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bridging Jensen Gap for Max-Min Group Fairness Optimization in Recommendation." pith.science (2026). https://pith.science/paper/6NLWO2WF

@misc{pith2026250209319,
  author       = {Pith},
  title        = {Pith review of: Bridging Jensen Gap for Max-Min Group Fairness Optimization in Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6NLWO2WF}},
  note         = {Machine review of arXiv:2502.09319}
}
read the original abstract

Group max-min fairness (MMF) is commonly used in fairness-aware recommender systems (RS) as an optimization objective, as it aims to protect marginalized item groups and ensures a fair competition platform. However, our theoretical analysis indicates that integrating MMF constraint violates the assumption of sample independence during optimization, causing the loss function to deviate from linear additivity. Such nonlinearity property introduces the Jensen gap between the model's convergence point and the optimal point if mini-batch sampling is applied. Both theoretical and empirical studies show that as the mini-batch size decreases and the group size increases, the Jensen gap will widen accordingly. Some methods using heuristic re-weighting or debiasing strategies have the potential to bridge the Jensen gap. However, they either lack theoretical guarantees or suffer from heavy computational costs. To overcome these limitations, we first theoretically demonstrate that the MMF-constrained objective can be essentially reformulated as a group-weighted optimization objective. Then we present an efficient and effective algorithm named FairDual, which utilizes a dual optimization technique to minimize the Jensen gap. Our theoretical analysis demonstrates that FairDual can achieve a sub-linear convergence rate to the globally optimal solution and the Jensen gap can be well bounded under a mini-batch sampling strategy with random shuffle. Extensive experiments conducted using six large-scale RS backbone models on three publicly available datasets demonstrate that FairDual outperforms all baselines in terms of both accuracy and fairness. Our data and codes are shared at https://github.com/XuChen0427/FairDual.

Figures

Figures reproduced from arXiv: 2502.09319 by the authors.

Figure 1
Figure 1. Loss converges simulation with 1000 users and 1000 items. Sub-figure (a) and (b) [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Overall workflow of FairDual under every two batches [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Sub-figure (a) conducts a simulation to show Jensen gap [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Sub-figure (a) and (b) describe the NDCG and MMF changes [PITH_FULL_IMAGE:figures/full_fig_p024_4.png]
Figure 5
Figure 5. Figure 5: Sub-figure (a) and (b) describe the NDCG and MMF changes [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 6
Figure 6. Figure 6: Sub-figure (a) illustrates the density distribution of embeddings for each hidden layer of [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A Reproducibility Study of Product-side Fairness in Bundle Recommendation

    cs.IR 2025-07 conditional novelty 6.0 of 10

    The first evaluation of product-side fairness in bundle recommendation shows bundle-level and item-level fairness diverge, and user preference for bundles versus items changes fairness outcomes.

Reference graph

Works this paper leans on

72 extracted references · 52 canonical work pages · cited by 1 Pith paper

  1. [1]

    Abdollahpouri and R

    H. Abdollahpouri and R. Burke. Multi-stakeholder recommendation and its connection to multi-sided fairness. arXiv preprint arXiv:1907.13158, 2019

    arXiv 1907

  2. [2]

    Abdollahpouri, G

    H. Abdollahpouri, G. Adomavicius, R. Burke, I. Guy, D. Jannach, T. Kamishima, J. Krasnodebski, and L. Pizzato. Multistakeholder recommendation: Survey and research directions. User Modeling and User-Adapted Interaction, 30 0 (1): 0 127--158, 2020

    work page 2020

  3. [3]

    J. D. Abernethy, P. Awasthi, M. Kleindessner, J. Morgenstern, C. Russell, and J. Zhang. Active sampling for min-max fairness. In International Conference on Machine Learning, pages 53--65. PMLR, 2022

    work page 2022

  4. [4]

    Agarwal, A

    A. Agarwal, A. Beygelzimer, M. Dud \' k, J. Langford, and H. Wallach. A reductions approach to fair classification. In International conference on machine learning, pages 60--69. PMLR, 2018

    work page 2018

  5. [5]

    Balseiro, H

    S. Balseiro, H. Lu, and V. Mirrokni. Regularized online allocation problems: Fairness and beyond. In International Conference on Machine Learning, pages 630--639. PMLR, 2021

    work page 2021

  6. [6]

    K. Bao, J. Zhang, W. Wang, Y. Zhang, Z. Yang, Y. Luo, F. Feng, X. He, and Q. Tian. A bi-step grounding paradigm for large language models in recommendation systems. arXiv preprint arXiv:2308.08434, 2023

    arXiv 2023

  7. [7]

    Beltagy, M

    I. Beltagy, M. E. Peters, and A. Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020

    arXiv 2004

  8. [8]

    R. I. Bo t , S.-M. Grad, and G. Wanka. On strong and total lagrange duality for convex optimization problems. Journal of Mathematical Analysis and Applications, 337 0 (2): 0 1315--1325, 2008

    work page 2008

Show all 72 references
  1. [9]

    F. P. Calmon, D. Wei, B. Vinzamuri, K. N. Ramamurthy, and K. R. Varshney. Optimized pre-processing for discrimination prevention. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, page 3995–4004, Red Hook, NY, USA, 2017. Cur...

    2017

  2. [10]

    Chai and X

    J. Chai and X. Wang. Fairness with adaptive weights. In International Conference on Machine Learning, pages 2853--2866. PMLR, 2022

    2022

  3. [11]

    X. Chen, W. Fan, J. Chen, H. Liu, Z. Liu, Z. Zhang, and Q. Li. Fairly adaptive negative sampling for recommendations. In Proceedings of the ACM Web Conference 2023, pages 3723--3733, 2023 a

    2023

  4. [12]

    X. Chen, W. Fan, J. Chen, H. Liu, Z. Liu, Z. Zhang, and Q. Li. Fairly adaptive negative sampling for recommendations. In Proceedings of the ACM Web Conference 2023, pages 3723--3733, 2023 b

    2023

  5. [13]

    C. W. Churchman, R. L. Ackoff, and E. L. Arnoff. Introduction to operations research. 1957

    1957

  6. [14]

    C. Cousins. An axiomatic theory of provably-fair welfare-centric machine learning. Advances in Neural Information Processing Systems, 34: 0 16610--16621, 2021

    2021

  7. [15]

    C. Cousins. Uncertainty and the social planner’s problem: Why sample complexity matters. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 2004--2015, 2022

    2022

  8. [16]

    C. Cousins. Revisiting fair-pac learning and the axioms of cardinal welfare. In International Conference on Artificial Intelligence and Statistics, pages 6422--6442. PMLR, 2023

    2023

  9. [17]

    S. Dai, N. Shao, H. Zhao, W. Yu, Z. Si, C. Xu, Z. Sun, X. Zhang, and J. Xu. Uncovering chatgpt’s capabilities in recommender systems. In Proceedings of the 17th ACM Conference on Recommender Systems, RecSys '23, page 1126–1132, New York, NY, USA, 2023. Association for Computin...

    2023

  10. [18]

    Demidovich, G

    Y. Demidovich, G. Malinovsky, I. Sokolov, and P. Richt \'a rik. A guide through the zoo of biased sgd. Advances in Neural Information Processing Systems, 36: 0 23158--23171, 2023

    2023

  11. [19]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018

    2018 arXiv

  12. [20]

    Diamond and S

    S. Diamond and S. Boyd. Cvxpy: A python-embedded modeling language for convex optimization. The Journal of Machine Learning Research, 17 0 (1): 0 2909--2913, 2016

    2016

  13. [21]

    Do and N

    V. Do and N. Usunier. Optimizing generalized gini indices for fairness in rankings. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 737--747, 2022

    2022

  14. [22]

    V. Do, S. Corbett-Davies, J. Atif, and N. Usunier. Two-sided fairness in rankings via lorenz dominance. Advances in Neural Information Processing Systems, 34: 0 8596--8608, 2021

    2021

  15. [23]

    Dr \`e ze and N

    J. Dr \`e ze and N. Stern. Policy reform, shadow prices, and market prices. Journal of public economics, 42 0 (1): 0 1--45, 1990

    1990

  16. [24]

    J. Fan, Z. Wang, Y. Xie, and Z. Yang. A theoretical analysis of deep q-learning. In Learning for dynamics and control, pages 486--489. PMLR, 2020

    2020

  17. [25]

    X. Gao, M. Sitharam, and A. E. Roitberg. Bounds on the jensen gap, and implications for mean-concentrated distributions. arXiv preprint arXiv:1712.05267, 2017

    2017 arXiv

  18. [26]

    Hashimoto, M

    T. Hashimoto, M. Srivastava, H. Namkoong, and P. Liang. Fairness without demographics in repeated loss minimization. In International Conference on Machine Learning, pages 1929--1938. PMLR, 2018

    1929

  19. [27]

    He and J

    R. He and J. McAuley. Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering. In proceedings of the 25th international conference on world wide web, pages 507--517, 2016

    2016

  20. [28]

    X. He, H. Zhang, M.-Y. Kan, and T.-S. Chua. Fast matrix factorization for online recommendation with implicit feedback. In Proceedings of the 39th International ACM SIGIR conference on Research and Development in Information Retrieval, pages 549--558, 2016

    2016

  21. [29]

    X. He, L. Liao, H. Zhang, L. Nie, X. Hu, and T.-S. Chua. Neural collaborative filtering. In Proceedings of the 26th international conference on world wide web, pages 173--182, 2017

    2017

  22. [30]

    Hidasi, A

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

    2015 arXiv

  23. [31]

    E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021

    2021 arXiv

  24. [32]

    Y. Hu, S. Zhang, X. Chen, and N. He. Biased stochastic first-order methods for conditional stochastic optimization and applications in meta learning. Advances in Neural Information Processing Systems, 33: 0 2759--2770, 2020

    2020

  25. [33]

    Z. Hu, Y. Xu, and X. Tian. Adaptive priority reweighing for generalizing fairness improvement. In 2023 International Joint Conference on Neural Networks (IJCNN), pages 01--08. IEEE, 2023

    2023

  26. [34]

    Jiang, K

    M. Jiang, K. Bao, J. Zhang, W. Wang, Z. Yang, F. Feng, and X. He. Item-side fairness of large language model-based recommendation system, 2024

    2024

  27. [35]

    Jones, B

    A. Jones, B. Sufrin, and et al. EU competition law: text, cases, and materials. Oxford University Press, USA, 2014

    2014

  28. [36]

    Kamishima and S

    T. Kamishima and S. Akaho. Considerations on recommendation independence for a find-good-items task. 2017

    2017

  29. [37]

    Kang and J

    W.-C. Kang and J. McAuley. Self-attentive sequential recommendation. In 2018 IEEE international conference on data mining (ICDM), pages 197--206. IEEE, 2018 a

    2018

  30. [38]

    Kang and J

    W.-C. Kang and J. McAuley. Self-attentive sequential recommendation. In 2018 IEEE international conference on data mining (ICDM), pages 197--206. IEEE, 2018 b

    2018

  31. [39]

    D. Kozlov. Combinatorial algebraic topology, volume 21. Springer Science & Business Media, 2008

    2008

  32. [40]

    Lan and M

    T. Lan and M. Chiang. An axiomatic theory of fairness in resource allocation. George Washington University, http://www. seas. gwu. edu/tlan/papers/fairness. pdf, Tech. Rep, 2011

    2011

  33. [41]

    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, KDD '23, page 1258–1267, New York, N...

    2023

  34. [42]

    Y. Li, H. Chen, S. Xu, Y. Ge, and Y. Zhang. Towards personalized fairness based on causal notion. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1054--1063, 2021

    2021

  35. [43]

    X. Lin, W. Wang, Y. Li, F. Feng, S.-K. Ng, and T.-S. Chua. Bridging items and language: A transition paradigm for large language model-based recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1816--1826. ACM, 2024 a

    2024

  36. [44]

    X. Lin, W. Wang, Y. Li, S. Yang, F. Feng, Y. Wei, and T.-S. Chua. Data-efficient fine-tuning for llm-based recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 365--374. ACM, 2024 b

    2024

  37. [45]

    Lindenstrauss, G

    J. Lindenstrauss, G. Olsen, and Y. Sternfeld. The poulsen simplex. In Annales de l'institut Fourier, volume 28, pages 91--114, 1978

    1978

  38. [46]

    Marras, L

    M. Marras, L. Boratto, G. Ramos, and G. Fenu. Equality of learning opportunity via individual fairness in personalized recommendations. International Journal of Artificial Intelligence in Education, 32 0 (3): 0 636--684, 2022

    2022

  39. [47]

    implicit

    D. Matten, J. Moon, and et al. “implicit” and “explicit” csr: A conceptual framework for a comparative understanding of corporate social responsibility. Academy of management Review, 33 0 (2): 0 404--424, 2008

    2008

  40. [48]

    Naghiaei, H

    M. Naghiaei, H. A. Rahmani, and Y. Deldjoo. Cpfair: Personalized consumer and producer fairness re-ranking for recommender systems. arXiv preprint arXiv:2204.08085, 2022

    2022 arXiv

  41. [49]

    Narasimhan, A

    H. Narasimhan, A. Cotter, M. Gupta, and S. Wang. Pairwise fairness for ranking and regression. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5248--5255, 2020

    2020

  42. [50]

    Paszke, S

    A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer. Automatic differentiation in pytorch. 2017

    2017

  43. [51]

    G. K. Patro, A. Biswas, N. Ganguly, K. P. Gummadi, and A. Chakraborty. Fairrec: Two-sided fairness for personalized recommendations in two-sided platforms. In Proceedings of The Web Conference 2020, pages 1194--1204, 2020

    2020

  44. [52]

    Rendle, C

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

    2012 arXiv

  45. [53]

    Y. Roh, K. Lee, S. E. Whang, and C. Suh. Fairbatch: Batch selection for model fairness. arXiv preprint arXiv:2012.01696, 2020

    2012 arXiv

  46. [54]

    A. P. Singh and G. J. Gordon. A unified view of matrix factorization models. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 358--373. Springer, 2008

    2008

  47. [55]

    F. Sun, J. Liu, J. Wu, C. Pei, X. Lin, W. Ou, and P. Jiang. Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM international conference on information and knowledge management, pages 1441--1450, 2019

    2019

  48. [56]

    R. Sun. Optimization for deep learning: theory and algorithms. arXiv preprint arXiv:1912.08957, 2019

    1912 arXiv

  49. [57]

    J. Tang, S. Shen, Z. Wang, Z. Gong, J. Zhang, and X. Chen. When fairness meets bias: a debiased framework for fairness aware top-n recommendation. In Proceedings of the 17th ACM Conference on Recommender Systems, RecSys '23, page 200–210, New York, NY, USA, 2023. Association f...

    2023

  50. [58]

    Touvron, L

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y. Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023

    2023 arXiv

  51. [59]

    Ullah, M

    H. Ullah, M. Adil Khan, and T. Saeed. Determination of bounds for the jensen gap and its applications. Mathematics, 9 0 (23): 0 3132, 2021

    2021

  52. [60]

    Van der Maaten and G

    L. Van der Maaten and G. Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 0 (11), 2008

    2008

  53. [61]

    H. Wen, X. Yi, T. Yao, J. Tang, L. Hong, and E. H. Chi. Distributionally-robust recommendations for improving worst-case user experience. In Proceedings of the ACM Web Conference 2022, pages 3606--3610, 2022

    2022

  54. [62]

    C. Wu, F. Wu, S. Ge, T. Qi, Y. Huang, and X. Xie. Neural news recommendation with multi-head self-attention. In K. Inui, J. Jiang, V. Ng, and X. Wan, editors, Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint...

    2019 doi

  55. [63]

    F. Wu, Y. Qiao, J.-H. Chen, C. Wu, T. Qi, J. Lian, D. Liu, X. Xie, J. Gao, W. Wu, et al. Mind: A large-scale dataset for news recommendation. In ACL, pages 3597--3606, 2020

    2020

  56. [64]

    Y. Wu, J. Cao, G. Xu, and Y. Tan. Tfrom: A two-sided fairness-aware recommendation model for both customers and providers. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1013--1022, 2021

    2021

  57. [65]

    Xiong, N

    Z. Xiong, N. Dalmasso, A. Mishler, V. K. Potluru, T. Balch, and M. Veloso. Fairwasp: Fast and optimal fair wasserstein pre-processing. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 16120--16128, 2024

    2024

  58. [66]

    C. Xu, J. Xu, X. Chen, Z. Dong, and J.-R. Wen. Dually enhanced propensity score estimation in sequential recommendation. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, pages 2260--2269, 2022

    2022

  59. [67]

    C. Xu, S. Chen, J. Xu, W. Shen, X. Zhang, G. Wang, and Z. Dong. P-mmf: Provider max-min fairness re-ranking in recommender system. In Proceedings of the ACM Web Conference 2023, pages 3701--3711, 2023

    2023

  60. [68]

    C. Xu, W. Wang, Y. Li, L. Pang, J. Xu, and T.-S. Chua. A study of implicit ranking unfairness in large language models. In Y. Al-Onaizan, M. Bansal, and Y.-N. Chen, editors, Findings of the Association for Computational Linguistics: EMNLP 2024, pages 7957--7970, Miami, Florida...

    2024 doi

  61. [69]

    C. Xu, J. Xu, Y. Ding, X. Zhang, and Q. Qi. Fairsync: Ensuring amortized group exposure in distributed recommendation retrieval, 2024 b

    2024

  62. [70]

    C. Xu, X. Ye, W. Wang, L. Pang, J. Xu, and T.-S. Chua. A taxation perspective for fair re-ranking. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '24, page 1494–1503, New York, NY, USA, 2024 c . Associa...

    2024

  63. [71]

    Yao and B

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

    2017

  64. [72]

    Z. Zhu, J. Wang, and J. Caverlee. Measuring and mitigating item under-recommendation bias in personalized ranking systems. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '20, page 449–458, New York, NY,...

    2020

Pith tools

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