Pith. sign in

REVIEW 2 major objections 7 minor 1 cited by

User-item fairness tradeoffs in recommendations

T0 review · 2 major / 7 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Item-fairness constraints on recommendations cost users little when preferences are diverse, yet can push users with misestimated tastes toward the most unpopular items.

desk verdict Solid theory on how diversity lowers the price of item fairness, but the 'free fairness' headline overstates what Theorem 3 proves. read the letter →

arxiv 2412.04466 v1 pith:RQQ5MGL4 submitted 2024-12-05 cs.IR cs.CY

Sophie Greenwood , Sudalakshmee Chiniah , Nikhil Garg This is my paper · ORCID
classification cs.IRcs.CY MSC 90C0590C47
keywords recommendersystemsitemfairnessuserpriceofmulti-sidedpreferencediversitymisestimationcold-startusers
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 asks a concrete question: when a recommender system tries to be fair to both users and items at once, how much must users give up, and which users pay? It models the platform as maximizing the worst-off user's normalized utility subject to a guarantee that the worst-off item receives at least a fraction $\gamma$ of the exposure it could get, and it characterizes the solutions of this constrained optimization. Two phenomena emerge. 'Free fairness': when user preferences are diverse, the item-fairness constraint can be tightened at almost no cost to users, while homogeneous populations face a steep tradeoff. 'Reinforced disparate effects': users whose preferences the platform estimates poorly (for example, cold-start users treated as average) are steered toward globally unpopular items under item fairness, and in the worst case this misestimation cost can be made arbitrarily large even though the platform is explicitly maximizing the minimum user utility. A prototype recommender for newly posted papers on a large preprint server confirms the diversity prediction and shows that the misestimation price is high in practice, though on average it is not worsened by item fairness.

What carries the argument

The load-bearing object is a constrained max-min program: maximize the minimum normalized user utility $U_{\min}(\rho) = \min_i (\sum_j \rho_{ij} w_{ij} / \max_j w_{ij})$ subject to the minimum normalized item utility $I_{\min}(\rho) \geq \gamma I^{*}_{\min}$, where $I^{*}_{\min}$ itself is the value of another max-min problem. Because that feasible region is defined by a nested optimization, the paper's workhorse (Propositions 1 and 2) reduces the problem to a linear program whenever the optimum lies in a linearly describable set of policies, and proves a sparsity property inside the set of policies that treat identical user types identically: at most $n + K - 1$ type–item pairs receive positive probability, and at most $K - 1$ items are shared between types. For the two-type, opposite-preferences model this sparsity forces a unique closed-form solution governed by the ratios $q_j(\alpha) = \alpha v_j / (\alpha v_j + (1-\alpha) v_{n-j+1})$, whose monotonicity in $\alpha$ and $j$ drives Theorem 3. Theorem 4 adds a column-symmetric policy set $S'$ for the three-type misestimation model; its pivot structure is what shows that misestimated users never receive their favorite item whenever $\beta > 1/n$.

What would settle it

A measurement that would settle the question: in a platform that logs user-side and item-side gains separately (for instance, user engagement time versus creator revenue per recommendation), compute the price-of-fairness curve for user segments ordered by measured preference diversity. The free-fairness claim predicts the curve is systematically flatter for high-diversity segments and steeper near maximal item fairness for homogeneous ones; the misestimation claim predicts that the lowest-confidence users are steered toward the globally least-preferred items as $\gamma$ approaches 1. Either pattern failing to appear — or the curves inverting when item utilities are separable but misaligned with user utilities — would falsify the paper's characterization.

Watch

Extended reading notes

Core claim

The central discovery is that the user–item fairness tradeoff is not a fixed constant but a function of the population's preference structure and of the platform's estimation error. Theorem 3 shows that in a population with two user types holding opposite preferences, with $\alpha$ the share of one type, the price of fairness — the relative drop in minimum normalized user utility when item fairness is pushed to its maximum — is minimized at $\alpha = 1/2$: it decreases in $\alpha$ on $(0, 1/2]$ and increases on $[1/2, 1)$. A mixed, diverse population thus obtains item fairness 'for free,' whereas a homogeneous population pays a near-linear cost, as in the two-item example where guaranteeing a minority item even $\epsilon$ of the users' utility forces user utility down by roughly the same $\epsilon$. Theorem 4 adds misestimation: when a fraction $1-\beta$ of users are 'new' and the platform estimates each of their utilities as the average of the two known types, then as long as $\beta > 1/n$ the price of misestimation under maximal item fairness can be driven above $1-\epsilon$ for any $\epsilon$ — those users end up being recommended the globally least preferred items — while without item fairness the same price is at most $1/2$. The empirical prototype bears out the diversity prediction: clusters of homogeneous users show steeper tradeoff curves, and the price of misestimation is already large even though, on average, item fairness does not add to it.

Load-bearing premise

The load-bearing premise is that the same number $w_{ij}$ measures both what user $i$ gains and what item $j$ gains from a recommendation; if items value exposure differently from how users value relevance in a way that rescaling cannot capture, the tradeoff characterizations in Theorems 3 and 4 — and the 'free fairness' conclusion — can fail, as the authors acknowledge.

Editorial extensions

If this is right

  • A platform that imposes item fairness on the whole user population gets it much more cheaply than imposing the same constraint on each homogeneous sub-market separately, so sub-market-level fairness constraints should be avoided.
  • Adding a user-fairness objective does not mechanically protect the users the platform knows least: because the optimization runs on the same misestimated utilities, the worst-off estimated users are precisely the ones handed the globally least-preferred items.
  • Tradeoff curves have a characteristic shape — roughly flat for moderate item-fairness guarantees and steep only near maximal item fairness ($\gamma \to 1$ in the empirical data) — so most of the item-fairness benefit is obtainable at low user cost, with sharp deterioration at the margin.
  • The empirical price of misestimation is already high without item fairness, and on average fairness constraints do not raise it; instance-specific measurement, rather than blanket statements, is required to predict the effect in a given setting.

Reading between the lines

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

  • A testable extension of Theorem 4: in a deployed recommender that logs confidence intervals for user-taste estimates, plot the minimum normalized utility of low-confidence users against $\gamma$; the theorem predicts a downward jump in their relative utility as $\gamma$ approaches 1, while Theorem 3 predicts the whole curve flattens as the user base's preference diversity grows.
  • The framework doubles as a design diagnostic: compute the price-of-fairness curve on holdout data separately for high- and low-diversity user segments before deploying any fairness constraint — a diverse user base makes item fairness nearly free, while a niche platform with one dominant taste should expect real user cost.
  • The single-item slate assumption may understate the free-fairness result: with multi-item slates a platform could satisfy item exposure within a slate, and the cold-start harm of Theorem 4 could shrink because the least-preferred item occupies only one slot among several.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 7 minor

Summary. The paper studies user-item fairness tradeoffs in a max-min recommendation model with symmetric user-item utilities. It defines the price of fairness as the relative loss in minimum normalized user utility when maximal item fairness is imposed, and the price of misestimation as the additional loss from optimizing with estimated utilities. The main theoretical results are Theorem 3, showing in a two-type opposite-preference model that the price of fairness is minimized at balanced population shares, and Theorem 4, showing that with cold-start users whose preferences are averaged, item fairness constraints can arbitrarily worsen the price of misestimation when the known population share exceeds 1/n. The paper also prototypes an arXiv preprint recommender using text-embedding similarities as utilities and reports empirical tradeoff curves for homogeneous versus diverse users and for misestimated users.

Significance. The theoretical framework is a genuine contribution: Proposition 1 reduces the item-fair constrained max-min problem to a linear program under a uniqueness condition, Proposition 2 establishes a sparsity structure, and the appendix proofs are detailed and internally plausible. The paper ships reproducible code, uses no fitted parameters in the theoretical model, and makes the context-dependence of fairness tradeoffs concrete. If the claims are appropriately scoped, the work is a useful conceptual guide for when item-fairness constraints are cheap or costly and for where the burden of misestimation falls. However, the headline “free fairness” is stronger than Theorem 3 supports, and the empirical validation of the utility proxy is weak, so the significance is contingent on revision.

major comments (2)
  1. [Abstract and Section 4, Theorem 3] The abstract and introduction claim that when user preferences are diverse there is “free” item and user fairness, but Theorem 3 only proves that the price of fairness is minimized at α = 1/2 in the two-type, opposite-preferences model; it gives no bound on the value of that minimum. Evaluating the paper’s own closed form in Lemma 5 at α = 1/2 with n = 5 and v = (5,4,3,2,1) gives U*_min(1, 1/2) = 30/37 ≈ 0.811, so the price of fairness is approximately 0.189 even at maximal diversity, and for linear value profiles this price grows with n. Figure 3a likewise shows a steep tradeoff as γ → 1 even for heterogeneous users. The abstract and introduction should be revised to claim that diversity reduces the price of fairness, not that fairness is free.
  2. [Section 6, Table 1] The empirical demonstration of the phenomena rests on text-embedding similarity scores as user-item utilities, and the only validation of this proxy is the logistic regression in Table 1, which reports adjusted R^2 values between 0.046 and 0.135 for predicting future citations. Because these scores explain only a small fraction of the variation in the outcome they are meant to represent, the curves in Figure 3 should be presented as an illustration of the model on a plausibly relevant proxy rather than as a measurement of the phenomena in a real market. The authors should either provide additional validation of the utility proxy (for example, calibration against revealed engagement data) or substantially temper the empirical claims.
minor comments (7)
  1. [References] Reference [12] is missing its title, and reference [26] is listed as “Private correspondence”; both should be replaced with complete, publicly citable versions.
  2. [Section 6.1, Figure references] The text refers to “Figure 3a” and “Figure 3b” when discussing the main empirical results, but the corresponding figures in the main text are Figure 1(a) and Figure 1(b); the numbering should be corrected.
  3. [Section 6.1] The sentence “we compute U*_min(1, \w)” for 50 values of γ should read U*_min(γ, \w), since γ is the parameter being varied.
  4. [Appendix E, proof of Lemma 13] The displayed simplification “2βq_i (1−2δ)/(2βq_i) z_i” appears to contain an algebraic typo; the preceding line should simplify to 2βq_iδ, which is still positive by the definition of δ.
  5. [Throughout] There are several typographical errors, including “prepints” in the introduction, “in constrast” in Section 7, and “It is would be interesting” in Section 8.
  6. [Table 1] The table reports “Adjusted R2” for logistic regressions; because ordinary least-squares R^2 is not defined for logistic models, the authors should state which pseudo-R^2 measure is being used.
  7. [Section 2 and Appendix A] Theorems 3 and 4 are proved only under the symmetric shared-utility assumption, and the discussion in Section 8 correctly notes that with misaligned utilities the price of fairness can be arbitrarily high. The abstract and introduction should carry this caveat when stating the phenomena, since Appendix A only covers separable scaling or an empirical interpolation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the theorems are derived from a stated optimization model with explicit inputs; no fitted parameter or self-citation chain forces the claimed results.

full rationale

The paper's formal results, Theorem 3 and Theorem 4, are derived from the LP reduction in Propositions 1 and 2, using the stated two-type and three-type utility structures in Appendices D and E. The quantities alpha, beta, and v_j are model inputs, not parameters fitted to the conclusions; the closed forms in Lemmas 5 and 15 are solved from the linear program, not assumed. The symmetric shared-utility assumption w_ij is explicitly stated in Section 2 as a modeling choice and is relaxed in Appendix A; it is not defined in terms of the price of fairness, and the proofs do not use the target result as an input. The empirical sections compute U*_min(gamma, w) by solving the same convex optimization on real similarity scores; this measures the phenomena rather than fitting theory to data, and the logistic-regression validation of the similarity scores is independent evidence that the utility estimates carry signal. Self-citations, such as [20] and [41] for the max-score choice and [40] for motivation, are not load-bearing for the theorems, and no uniqueness result is imported from the authors' prior work. A claim-support gap may exist, since Theorem 3 proves a comparative monotonicity statement and does not bound the value of the price of fairness at alpha = 1/2, but that is a strength/correctness concern, not circularity. The paper also transparently states its limitations in Section 8, including the restricted two- and three-type settings and the symmetric-utility assumption, and those admissions reinforce rather than undermine the non-circular nature of the derivation. No circular step can be exhibited as a specific equation-to-equation reduction, so the appropriate score is 0.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central claims rest on the max-min fairness model and its normalization choices, not on fitted parameters. The empirical study adds design parameters (k=41 clusters, 10% misestimated users) that affect the measurement but not the theoretical results.

free parameters (2)
  • Number of k-means clusters (k=41) = 41
    Chosen by the authors so each cluster contains about 500 users in expectation; defines 'homogeneous' user groups in the empirical study. Not fitted to the outcome, but a hand-chosen experimental design parameter.
  • Misestimation fraction (10%) = 0.10
    Fraction of users whose preferences are replaced by the average of the other users in Figure 3b; a design choice for the empirical experiment.
assumptions (6)
  • domain assumption Users receive exactly one recommendation (unit demand).
    Used in the formal model (Section 2) and throughout; the paper notes in Section 8 that the framework does not extend to multiple recommendations.
  • domain assumption User and item utilities are symmetric: wij is both the user's utility and the item's utility for a recommendation.
    Stated in Section 2 and relied on by Theorems 3 and 4; the paper argues in Appendix A that scaling by ai, bj cancels in normalization, but full misalignment is not theoretically handled.
  • domain assumption Fairness is measured by the minimum normalized utility (max-min / egalitarian fairness).
    Stated in Section 2; the paper says results might not extend to other definitions, and only empirical extensions are given in Appendix A.
  • domain assumption Item fairness is imposed as a lower bound on the minimum normalized item utility: Imin >= γ I*_min.
    Problem 1; the paper says this is conceptually dual to adding an item fairness objective term.
  • ad hoc to paper Theorem 3 assumes two user types with opposite preference orders.
    This is the specific population model used to prove the diversity result; the framework is more general but the theorem is restricted.
  • ad hoc to paper Theorem 4 assumes misestimated users' utilities are the average of the two opposing type vectors.
    This is the specific misestimation model used in the worst-case construction; real misestimation could differ.

how reviews work

0 comments
Cite this review

Pith. "Pith review of User-item fairness tradeoffs in recommendations." pith.science (2026). https://pith.science/paper/RQQ5MGL4

@misc{pith2026241204466,
  author       = {Pith},
  title        = {Pith review of: User-item fairness tradeoffs in recommendations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RQQ5MGL4}},
  note         = {Machine review of arXiv:2412.04466}
}
read the original abstract

In the basic recommendation paradigm, the most (predicted) relevant item is recommended to each user. This may result in some items receiving lower exposure than they "should"; to counter this, several algorithmic approaches have been developed to ensure item fairness. These approaches necessarily degrade recommendations for some users to improve outcomes for items, leading to user fairness concerns. In turn, a recent line of work has focused on developing algorithms for multi-sided fairness, to jointly optimize user fairness, item fairness, and overall recommendation quality. This induces the question: what is the tradeoff between these objectives, and what are the characteristics of (multi-objective) optimal solutions? Theoretically, we develop a model of recommendations with user and item fairness objectives and characterize the solutions of fairness-constrained optimization. We identify two phenomena: (a) when user preferences are diverse, there is "free" item and user fairness; and (b) users whose preferences are misestimated can be especially disadvantaged by item fairness constraints. Empirically, we prototype a recommendation system for preprints on arXiv and implement our framework, measuring the phenomena in practice and showing how these phenomena inform the design of markets with recommendation systems-intermediated matching.

Figures

Figures reproduced from arXiv: 2412.04466 by the authors.

Figure 1
Figure 1. Empirical (using our arXiv recommender) tradeoff between the minimum user (Y axis) and [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. We repeat the experiment of Figure [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. We repeat the experiment of Figure [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Robustness of empirical findings without symmetry assumption: user-item fairness trade [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Robustness of empirical findings without symmetry assumption: item fairness constraints [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Distribution of research paper publications over time [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Distribution of the number of papers per user in the training set on a logarithmic scale [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Pr(score|User cites paper), the distribution of the score for a user-paper pair, conditional on whether the user cites the paper in the future, for the Max score, TF-IDF model. Variable Coefficient Std. Err z-value P-value Adjusted R2 Similarity score 12.4100 0.058 212…
Figure 9
Figure 9. Figure 9: Pr(score|Author cites paper), the distribution of the score for a user-paper pair, conditional on whether the user cites the paper in the future, for the Mean score, TF-IDF model. Variable Coefficient Std. Err z-value P-value Adjusted R2 Similarity score 20.2122 0.131 …
Figure 10
Figure 10. Figure 10: Pr(score|Author cites paper), the distribution of the score for a user-paper pair, conditional on whether the user cites the paper in the future, for the Max score, Sentence transformer model. Max score, Sentence transformer [PITH_FULL_IMAGE:figures/full_fig_p023_10.png]
Figure 11
Figure 11. Figure 11: Pr(score|User cites paper), the distribution of the score for a user-paper pair, conditional on whether the user cites the paper in the future, for the Mean score, Sentence transformer model. Mean score, Sentence transformer [PITH_FULL_IMAGE:figures/full_fig_p023_11.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. FACTER: Fairness-Aware Conformal Thresholding and Prompt Engineering for Enabling Fair LLM-Based Recommender Systems

    cs.IR 2025-02 reject novelty 5.0 of 10

    FACTER combines conformal thresholding with iterative prompt editing to reduce demographic bias in black-box LLM recommendations, reporting up to 95.5% fewer fairness violations.

Reference graph

Works this paper leans on

70 extracted references · 49 canonical work pages · cited by 1 Pith paper

  1. [1]

    arXiv dataset, 2024

    arXiv.org submitters. arXiv dataset, 2024. URL https://www.kaggle.com/dsv/7548853

    arXiv 2024

  2. [2]

    An approximation algorithm for max-min fair allocation of indivisible goods

    Arash Asadpour and Amin Saberi. An approximation algorithm for max-min fair allocation of indivisible goods. In Proceedings of the Thirty-Ninth Annual ACM Symposium on Theory of Computing, STOC ’07, page 114–121, New York, NY , USA, 2007. Association for Computing Machinery. ISBN 9781595936318. doi: 10.1145/1250790.1250808. URL https://doi.org/ 10.1145/12...

    arXiv 2007

  3. [3]

    Assort- ment optimization with visibility constraints, 2024

    Theo Barre, Omar El Housni, Marouane Ibn Brahim, Andrea Lodi, and Danny Segev. Assort- ment optimization with visibility constraints, 2024. URL https://arxiv.org/abs/2307. 13656

    work page 2024

  4. [4]

    Mostly exploration-free algorithms for contextual bandits

    Hamsa Bastani, Mohsen Bayati, and Khashayar Khosravi. Mostly exploration-free algorithms for contextual bandits. Management Science, 67, 07 2020. doi: 10.1287/mnsc.2020.3605

    arXiv 2020

  5. [5]

    A framework for fairness in two-sided marketplaces, 2020

    Kinjal Basu, Cyrus DiCiccio, Heloise Logan, and Noureddine El Karoui. A framework for fairness in two-sided marketplaces, 2020. URL https://arxiv.org/abs/2006.12756

    arXiv 2020

  6. [6]

    The price of fairness for indivisible goods

    Xiaohui Bei, Xinhang Lu, Pasin Manurangsi, and Warut Suksompong. The price of fairness for indivisible goods. Theory of Computing Systems, 65:1069 – 1093, 2019. URL https: //api.semanticscholar.org/CorpusID:152282391

    work page 2019

  7. [7]

    Introduction to Linear Optimization

    Dimitris Bertsimas and John Tsitsiklis. Introduction to Linear Optimization. 1998

    work page 1998

  8. [8]

    Farias, and Nikolaos Trichakis

    Dimitris Bertsimas, Vivek F. Farias, and Nikolaos Trichakis. The price of fairness. Operational Research, 59:17–31, 2011

    work page 2011

Show all 70 references
  1. [9]

    The fault in our recommendations: On the perils of optimizing the measurable, 2024

    Omar Besbes, Yash Kanoria, and Akshit Kumar. The fault in our recommendations: On the perils of optimizing the measurable, 2024. URL https://arxiv.org/abs/2405.03948

    2024 arXiv

  2. [10]

    Multisided fairness for recommendation, 2017

    Robin Burke. Multisided fairness for recommendation, 2017. URL https://arxiv.org/ abs/1707.00093

    2017 arXiv

  3. [11]

    Balanced neighborhoods for multi-sided fairness in recommendation

    Robin Burke, Nasim Sonboli, and Aldo Ordonez-Gauger. Balanced neighborhoods for multi-sided fairness in recommendation. In Sorelle A. Friedler and Christo Wilson, editors, Proceedings of the 1st Conference on Fairness, Accountability and Transparency, volume 81 of Proceedings ...

    2018

  4. [12]

    URL https: //arxiv.org/abs/2306.10050

    Qinyi Chen, Jason Liang, Negin Golrezaei, and Djallel Bouneffouf, 2024. URL https: //arxiv.org/abs/2306.10050

    2024 arXiv

  5. [13]

    Arman Cohan, Sergey Feldman, Iz Beltagy, Doug Downey, and Daniel S. Weld. SPECTER: Document-level representation learning using citation-informed transformers. In ACL, 2020

    2020

  6. [14]

    Can probabilistic feedback drive user impacts in online platforms? In International Conference on Artificial Intelligence and Statistics, pages 2512–2520

    Jessica Dai, Bailey Flanigan, Meena Jagadeesan, Nika Haghtalab, and Chara Podimata. Can probabilistic feedback drive user impacts in online platforms? In International Conference on Artificial Intelligence and Statistics, pages 2512–2520. PMLR, 2024

    2024

  7. [15]

    Accounting for AI and users shaping one another: The role of mathematical models, 2024

    Sarah Dean, Evan Dong, Meena Jagadeesan, and Liu Leqi. Accounting for AI and users shaping one another: The role of mathematical models, 2024. URL https://arxiv.org/abs/2404. 12366. 11

    2024

  8. [16]

    Who pays? personalization, bossiness and the cost of fairness, 2022

    Paresha Farastu, Nicholas Mattei, and Robin Burke. Who pays? personalization, bossiness and the cost of fairness, 2022. URL https://arxiv.org/abs/2209.04043

    2022 arXiv

  9. [17]

    Assigning papers to referees

    Naveen Garg, Telikepalli Kavitha, Amit Kumar, Kurt Mehlhorn, and Julián Mestre. Assigning papers to referees. Algorithmica, 58(1):119–136, September 2010. ISSN 0178-4617

    2010

  10. [18]

    Designing optimal binary rating systems

    Nikhil Garg and Ramesh Johari. Designing optimal binary rating systems. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 1930–1939. PMLR, 2019

    1930

  11. [19]

    Designing informative rating systems: Evidence from an online labor market

    Nikhil Garg and Ramesh Johari. Designing informative rating systems: Evidence from an online labor market. Manufacturing & Service Operations Management, 23(3):589–605, 2021

    2021

  12. [20]

    The stereotyping problem in collaboratively filtered recommender systems

    Wenshuo Guo, Karl Krauth, Michael Jordan, and Nikhil Garg. The stereotyping problem in collaboratively filtered recommender systems. In Proceedings of the 1st ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization, pages 1–10, 2021

    2021

  13. [21]

    Recommending to strategic users,

    Andreas Haupt, Dylan Hadfield-Menell, and Chara Podimata. Recommending to strategic users,

  14. [22]

    Matching of users and creators in two-sided markets with departures, 2023

    Daniel Huttenlocher, Hannah Li, Liang Lyu, Asuman Ozdaglar, and James Siderius. Matching of users and creators in two-sided markets with departures, 2023. URL https://arxiv.org/ abs/2401.00313

    2023 arXiv

  15. [23]

    Supply-side equilibria in recommender systems

    Meena Jagadeesan, Nikhil Garg, and Jacob Steinhardt. Supply-side equilibria in recommender systems. Advances in Neural Information Processing Systems, 36, 2024

    2024

  16. [24]

    Price of fairness on networked auctions

    Mariusz Kaleta. Price of fairness on networked auctions. J. Appl. Math., 2014:860747:1– 860747:7, 2014. URL https://api.semanticscholar.org/CorpusID:38538799

    2014

  17. [25]

    Graham, F.Q

    Rodney Michael Kinney, Chloe Anastasiades, Russell Authur, Iz Beltagy, Jonathan Bragg, Alexandra Buraczynski, Isabel Cachola, Stefan Candra, Yoganand Chandrasekhar, Arman Cohan, Miles Crawford, Doug Downey, Jason Dunkelberger, Oren Etzioni, Rob Evans, Sergey Feldman, Joseph Go...

    2023 arXiv

  18. [26]

    Modeling amplification in a network of speakers and listeners

    Jon Kleinberg and Michela Meister. Modeling amplification in a network of speakers and listeners. Private correspondence, 2024

    2024

  19. [27]

    The challenge of understanding what users want: Inconsistent preferences and engagement optimization

    Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan. The challenge of understanding what users want: Inconsistent preferences and engagement optimization. Management Science, 2023

    2023

  20. [28]

    Mining of Massive Datasets

    Jure Leskovec, Anand Rajaraman, and Jeffrey David Ullman. Mining of Massive Datasets. Cambridge University Press, 2011

    2011

  21. [29]

    When collaborative filtering is not collabora- tive: Unfairness of PCA for recommendations, 2023

    David Liu, Jackie Baek, and Tina Eliassi-Rad. When collaborative filtering is not collabora- tive: Unfairness of PCA for recommendations, 2023. URL https://arxiv.org/abs/2310. 09687

    2023

  22. [30]

    Strategic ranking

    Lydia T Liu, Nikhil Garg, and Christian Borgs. Strategic ranking. In International Conference on Artificial Intelligence and Statistics, pages 2489–2518. PMLR, 2022

    2022

  23. [31]

    Personalizing fairness-aware re-ranking

    Weiwen Liu and Robin Burke. Personalizing fairness-aware re-ranking. In FATRecWorkshop on Responsible Recommendation, page to appear, 2018

    2018

  24. [32]

    Redesigning service level agreements: Equity and efficiency in city government operations

    Zhi Liu and Nikhil Garg. Redesigning service level agreements: Equity and efficiency in city government operations. In ACM Conference on Economics and Computation, 2024. 12

    2024

  25. [33]

    Balancing producer fairness and efficiency via Bayesian rating system design

    Thomas Ma, Michael S Bernstein, Ramesh Johari, and Nikhil Garg. Balancing producer fairness and efficiency via Bayesian rating system design. In International AAAI Conference on Web and Social Media, ICWSM ‘25, 2025

    2025

  26. [34]

    Redesigning V olunteer- Match’s ranking algorithm: Toward more equitable access to volunteers

    Vahideh Manshadi, Scott Rodilitz, Daniela Saban, and Akshaya Suresh. Redesigning V olunteer- Match’s ranking algorithm: Toward more equitable access to volunteers. Available at SSRN 4497747, 2023

    2023

  27. [35]

    Towards a fair marketplace: Counterfactual evaluation of the trade-off between relevance, fair- ness & satisfaction in recommendation systems

    Rishabh Mehrotra, James McInerney, Hugues Bouchard, Mounia Lalmas, and Fernando Diaz. Towards a fair marketplace: Counterfactual evaluation of the trade-off between relevance, fair- ness & satisfaction in recommendation systems. In Proceedings of the 27th ACM International Con...

    2018

  28. [36]

    Choosing the right weights: Balancing value, strategy, and noise in recommender systems, 2023

    Smitha Milli, Emma Pierson, and Nikhil Garg. Choosing the right weights: Balancing value, strategy, and noise in recommender systems, 2023. URL https://arxiv.org/abs/2305. 17428

    2023

  29. [37]

    Fair Division and Collective Welfare

    Hervé Moulin. Fair Division and Collective Welfare. MIT Press, 2003

    2003

  30. [38]

    Conic optimization via operator splitting and homogeneous self-dual embedding

    Brendan O’Donoghue, Eric Chu, Neal Parikh, and Stephen Boyd. Conic optimization via operator splitting and homogeneous self-dual embedding. Journal of Optimization Theory and Applications, 169(3):1042–1068, June 2016. URL http://stanford.edu/~boyd/papers/ scs.html

    2016

  31. [39]

    Gummadi, and Abhijnan Chakraborty

    Gourab K Patro, Arpita Biswas, Niloy Ganguly, Krishna P. Gummadi, and Abhijnan Chakraborty. FairRec: Two-sided fairness for personalized recommendations in two-sided platforms. In Proceedings of The Web Conference 2020, WWW ’20, page 1194–1204, New York, NY , USA,

    2020

  32. [40]

    Fair ranking: A critical review, challenges, and future directions

    Gourab K Patro, Lorenzo Porcaro, Laura Mitchell, Qiuyue Zhang, Meike Zehlike, and Nikhil Garg. Fair ranking: A critical review, challenges, and future directions. In Proceedings of the 2022 ACM conference on fairness, accountability, and transparency, pages 1929–1942, 2022. do...

    2022

  33. [41]

    Reconciling the accuracy-diversity trade-off in recommendations

    Kenny Peng, Manish Raghavan, Emma Pierson, Jon Kleinberg, and Nikhil Garg. Reconciling the accuracy-diversity trade-off in recommendations. In Proceedings of the ACM on Web Conference 2024, WWW ’24, page 1318–1329, New York, NY , USA, 2024. Association for Computing Machinery....

    2024

  34. [42]

    Rahmani, Yashar Deldjoo, Ali Tourani, and Mohammadmehdi Naghiaei

    Hossein A. Rahmani, Yashar Deldjoo, Ali Tourani, and Mohammadmehdi Naghiaei. The unfairness of active users and popularity bias in point-of-interest recommendation. In International Workshop on Algorithmic Bias in Search and Recommendation, 2022. doi: 10. 1007/978-3-031-09316-...

    2022 doi

  35. [43]

    A Theory of Justice: Original Edition

    John Rawls. A Theory of Justice: Original Edition. Harvard University Press, 1971. ISBN 9780674880108. URL http://www.jstor.org/stable/j.ctvjf9z6v

    1971

  36. [44]

    Fairness of exposure in rankings

    Ashudeep Singh and Thorsten Joachims. Fairness of exposure in rankings. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ’18, page 2219–2228, New York, NY , USA, 2018. Association for Computing Machinery. ISBN 97814503555...

    2018

  37. [45]

    Shah, and Aarti Singh

    Ivan Stelmakh, Nihar B. Shah, and Aarti Singh. Peerreview4all: Fair and accurate reviewer assignment in peer review. In Aurélien Garivier and Satyen Kale, editors, Proceedings of the 30th International Conference on Algorithmic Learning Theory, volume 98 of Proceedings of Mach...

    2019

  38. [46]

    Price of fairness in budget division for egalitarian social welfare

    Zhongzheng Tang, Chenhao Wang, and Mengqi Zhang. Price of fairness in budget division for egalitarian social welfare. In International Conference on Combinatorial Optimization and Applications, volume 12577 of Lecture Notes in Computer Science, pages 594–607. Springer, 2020. d...

    2020 doi

  39. [47]

    User fairness, item fairness, and diversity for rankings in two-sided markets

    Lequn Wang and Thorsten Joachims. User fairness, item fairness, and diversity for rankings in two-sided markets. In Proceedings of the 2021 ACM SIGIR International Conference on Theory of Information Retrieval, ICTIR ’21, page 23–41, New York, NY , USA, 2021. Associa- tion for...

    2021

  40. [48]

    Fairness of exposure in stochastic bandits

    Lequn Wang, Yiwei Bai, Wen Sun, and Thorsten Joachims. Fairness of exposure in stochastic bandits. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 10686...

    2021

  41. [49]

    Yong Xiao, Jianwei Huang, Chau Yuen, and Luiz A. Dasilva. Fairness and efficiency tradeoffs for user cooperation in distributed wireless networks. 2013 Proceedings IEEE INFOCOM, pages 285–289, 2013. doi: 10.1109/INFCOM.2013.6566780. URL https://ieeexplore. ieee.org/document/6566780

    2013

  42. [50]

    P-MMF: Provider max-min fairness re-ranking in recommender system

    Chen Xu, Sirui Chen, Jun Xu, Weiran Shen, Xiao Zhang, Gang Wang, and Zhenhua Dong. P-MMF: Provider max-min fairness re-ranking in recommender system. In Proceedings of the ACM Web Conference 2023, WWW ’23, page 3701–3711, New York, NY , USA, 2023. Association for Computing Mac...

    2023

  43. [51]

    market share

    Meike Zehlike, Francesco Bonchi, Carlos Castillo, Sara Hajian, Mohamed Megahed, and Ricardo Baeza-Yates. FA*IR: A fair top-k ranking algorithm. In Proceedings of the 2017 ACM Conference on Information and Knowledge Management, CIKM ’17, page 1569–1578, New York, NY , USA, 2017...

    2017

  44. [54]

    By contrast, t(α) is unique

    In particular, this occurs when for each j, xj = 0 or yj = 0. By contrast, t(α) is unique. 28 By Lemma 6, U ∗ min (α) = Ui(α) for i with τ (i) = 2. Then U ∗ min (1, α) = Ui(α) = 1 maxj vj nX j=1 yjvn−j+1 = 1 maxj vj  yt(α)vn−t(α)+1 + X j>t(α) yjvn−j+1   = 1 maxj vj    ...

  45. [55]

    Then a user i of type 1 will have normalized utility will be Ui(x, y) ≥ x1v1 v1 = I ∗ min q1 = I ∗ min v1 + vn vn ≥ I ∗ min > 1 n

    Likewise if n is odd, I ∗ min = 1 1 + qtRt + (1 − qt)Rt = 1 1 + Rt = 1 1 + P j>t 1 1−qj > 1 1 + 2(n − (n + 1)/2) = 1 n . Then a user i of type 1 will have normalized utility will be Ui(x, y) ≥ x1v1 v1 = I ∗ min q1 = I ∗ min v1 + vn vn ≥ I ∗ min > 1 n . Similarly a user of type...

  46. [56]

    Guidelines: • The answer NA means that the abstract and introduction do not include the claims made in the paper

    Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: We believe the abstract and the introduction provide an accurate overview of the contents of this paper. Guidelines: •...

  47. [57]

    Limitations

    Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: We discuss limitations in Section 8. Guidelines: • The answer NA means that the paper has no limitation while the answer No means that the paper has ...

  48. [58]

    We provide complete proofs of each theorem in Appendices C, D, and E

    Theory Assumptions and Proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [Yes] 47 Justification: For each of our main theorems, we provide a detailed description of the modeling assumpti...

  49. [59]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experimental Result Reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main ex- perimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and da...

  50. [60]

    Guidelines: • The answer NA means that paper does not include experiments requiring code

    Open access to data and code 48 Question: Does the paper provide open access to the data and code, with sufficient instruc- tions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: In the introduction, we ...

  51. [61]

    We describe the training procedure in Appendix B.2, and the evaluation in Appendix B.3

    Experimental Setting/Details Question: Does the paper specify all the training and test details (e.g., data splits, hyper- parameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: We provide a detailed descrip...

  52. [62]

    We also show error bars on the experimental plots and discuss how they were obtained

    Experiment Statistical Significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [Yes] Justification: In Appendix B we provide a thorough evaluation of ...

  53. [63]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experiments Compute Resources Question: For each experiment, does the paper provide sufficient information on the com- puter resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: We provide this informa...

  54. [64]

    • If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics

    Code Of Ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? Answer: [Yes] Justification: Guidelines: • The answer NA means that the authors have not reviewed the NeurIP...

  55. [65]

    Guidelines: • The answer NA means that there is no societal impact of the work performed

    Broader Impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [Yes] Justification: We discuss this in Section 8; since our paper is investigating potential negative social impacts of techn...

  56. [66]

    Guidelines: • The answer NA means that the paper poses no such risks

    Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? Answer: [NA] Justification: The paper poses...

  57. [67]

    Guidelines: • The answer NA means that the paper does not use existing assets

    Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: We cite the mode...

  58. [68]

    Guidelines: • The answer NA means that the paper does not release new assets

    New Assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [Yes] Justification: We release a code to train and analyze the recommendation system that we develop, and provide documentation with the...

  59. [69]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Crowdsourcing and Research with Human Subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? A...

  60. [70]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

  61. [2020]

    ISBN 9781450370233

    Association for Computing Machinery. ISBN 9781450370233. doi: 10.1145/3366423. 3380196. URL https://doi.org/10.1145/3366423.3380196

  62. [2023]

    URL https://arxiv.org/abs/2302.06559

    arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.