pith. sign in

arxiv: 2605.01524 · v1 · submitted 2026-05-02 · 💻 cs.IR

Post-hoc Provider Fairness Adaptation via Hierarchical Exposure Alignment

Jingzhi Li , Zhiyong Cheng , Richang Hong , Meng Wang This is my paper

Pith reviewed 2026-05-09 17:57 UTC · model grok-4.3

classification 💻 cs.IR
keywords provider fairnessrecommender systemspost-hoc adaptationexposure alignmenthierarchical fairnessfairness adapterNDCG optimizationexposure distribution
0
0 comments X

The pith

A post-hoc adapter attached to a frozen recommender can enforce provider fairness by learning additive score adjustments from embeddings and aligning exposure hierarchically.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to solve the problem that changing fairness goals in recommender systems usually requires either expensive retraining or inflexible post-hoc reranking. It introduces a lightweight fairness adapter that sits on top of an already-trained model and learns to tweak the original ranking scores using user-item embeddings. The adapter is trained by minimizing a KL divergence to a target fair exposure distribution, but the authors add hierarchical terms to also fix imbalances between provider groups and within groups. A joint differentiable NDCG loss is used at the same time so that accuracy does not suffer much. If successful, this would let platforms adjust provider exposure rules on demand without rebuilding the core ranking model each time.

Core claim

The paper claims that Post-hoc Fairness Adaptation (PFA) equips any frozen recommender with a fairness adapter that injects learned additive score adjustments into the original ranking scores. These adjustments are optimized by minimizing KL divergence to a target exposure distribution, augmented by Hierarchical Exposure Fairness Alignment (HEFA) that separately balances inter-group and intra-group disparities. The entire adapter is trained end-to-end together with a differentiable NDCG loss, allowing flexible fairness control while keeping ranking quality nearly intact.

What carries the argument

The Hierarchical Exposure Fairness Alignment (HEFA) mechanism, which augments global KL divergence minimization with explicit terms that correct both inter-group and intra-group provider exposure imbalances.

If this is right

  • Fairness targets can be changed after training without retraining the backbone recommender.
  • Exposure fairness is enforced at both the group level and the individual provider level rather than only in aggregate.
  • Ranking accuracy is maintained through simultaneous optimization with a differentiable NDCG loss.
  • The adapter outperforms standard post-hoc reranking methods on three public datasets while adding only negligible compute at inference time.
  • Diverse fairness requirements can be met by simply retraining the small adapter instead of the full model.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same adapter structure could be reused across different recommender backbones, reducing the cost of maintaining multiple fair versions of a system.
  • Hierarchical balancing terms might transfer to other ranking problems where imbalances exist at multiple scales, such as item categories or user cohorts.
  • In production, the adapter could be periodically updated on recent data to adapt to shifting provider populations without touching the core model.
  • One could test whether the approach still works when the target fairness distribution itself changes dynamically during live operation.

Load-bearing premise

That additive score adjustments learned from embeddings and optimized with KL divergence plus hierarchical terms plus joint NDCG loss will produce practically fair exposure distributions without creating new hidden imbalances.

What would settle it

Apply PFA to a new dataset, compute the global KL divergence, inter-group disparity, and intra-group disparity separately, and check whether the hierarchical terms reduce the two finer-grained disparities to near zero while NDCG remains within 1-2 percent of the original model.

Figures

Figures reproduced from arXiv: 2605.01524 by Jingzhi Li, Meng Wang, Richang Hong, Zhiyong Cheng.

Figure 1
Figure 1. Figure 1: Comparison of in-processing, post-processing, and view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed PFA framework. The frozen base recommender provides base scores, which are adjusted view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study across three datasets. (a) Component view at source ↗
Figure 5
Figure 5. Figure 5: Subgroup-level fairness comparison on Amazon view at source ↗
Figure 7
Figure 7. Figure 7: Sensitivity analysis of the HEFA weights view at source ↗
Figure 6
Figure 6. Figure 6: Sensitivity analysis of the accuracy-fairness trade view at source ↗
Figure 8
Figure 8. Figure 8: Inference time comparison across three datasets. view at source ↗
read the original abstract

Provider exposure fairness is crucial for sustaining a healthy content ecosystem and preventing monopolization in recommender systems. Yet, most existing methods either incorporate fairness constraints during model training, requiring expensive retraining when fairness objectives change, or rely on post-hoc reranking with fixed criteria, which lacks adaptability to diverse fairness requirements. To overcome these limitations, we propose Post-hoc Fairness Adaptation (PFA), a lightweight framework that equips a frozen recommender with a fairness adapter, enabling flexible fairness control without retraining the backbone model. Specifically, the fairness adapter learns personalized additive score adjustments from user-item embeddings, which are injected into the original ranking scores to steer provider exposure toward fairness. To train the adapter, we minimize the KL divergence between the actual and the target fair exposure distributions. However, this global objective implicitly treats all providers equally, ignoring structural disparities such as imbalanced provider group sizes and heterogeneous exposure within groups. Consequently, fairness may appear satisfied at an aggregate level while severe inter-group and intra-group exposure imbalances persist, undermining practical fairness. To address this, we design Hierarchical Exposure Fairness Alignment (HEFA), which explicitly balances inter- and intra-group provider exposure disparities, enabling flexible adaptation to diverse fairness requirements. To mitigate potential accuracy degradation, PFA jointly optimizes HEFA with a differentiable NDCG loss, enabling end-to-end fairness optimization while preserving ranking quality. Extensive experiments on three public datasets demonstrate that PFA achieves substantial fairness gains with negligible accuracy loss, consistently outperforming strong baselines.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes Post-hoc Fairness Adaptation (PFA), a lightweight post-hoc framework that attaches a fairness adapter to a frozen recommender model. The adapter learns personalized additive adjustments to ranking scores from user-item embeddings. These adjustments are optimized via Hierarchical Exposure Fairness Alignment (HEFA), which augments global KL divergence between actual and target exposure distributions with explicit inter-group and intra-group terms, and is jointly trained with a differentiable NDCG loss to preserve ranking quality. Experiments on three public datasets are reported to show substantial provider fairness gains with negligible accuracy loss relative to strong baselines.

Significance. If the central claims hold, the work provides a practical, retraining-free method for adapting provider exposure fairness to changing requirements, which addresses a clear limitation of both in-training constraint methods and fixed post-hoc rerankers. The hierarchical alignment component is a substantive technical contribution for handling group-size imbalances and intra-group heterogeneity that a purely global objective would miss. The joint NDCG optimization is a strength for maintaining utility.

major comments (2)
  1. [§3 (HEFA formulation)] §3 (HEFA formulation): The hierarchical KL terms are motivated as correcting inter- and intra-group imbalances that a global objective leaves unaddressed. However, the construction depends on the chosen group partition; if groups are defined at a coarse level (e.g., top-level categories), the intra-group term only equalizes within those buckets. The paper must show that the learned additive adjustments do not create or leave uncorrected exposure monopolization at finer partitions (e.g., by popularity strata or item metadata) that are not part of the hierarchy, as this directly affects whether the method delivers the claimed practical fairness.
  2. [§4 (Experiments)] §4 (Experiments): The headline result of 'substantial fairness gains with negligible accuracy loss' is supported by aggregate metrics. To substantiate that HEFA plus the joint NDCG term does not trade off against finer imbalances, the results must include (a) exposure variance or max-min ratios within each defined group and (b) an ablation or secondary evaluation on sub-partitions not used in the hierarchical loss. Without these, the claim that the method 'enables flexible adaptation to diverse fairness requirements' remains incompletely supported.
minor comments (1)
  1. [Abstract and §3] The abstract and method sections use 'provider group sizes' and 'heterogeneous exposure within groups' without an early, explicit definition or example of how groups are constructed from the datasets; adding this would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the constructive feedback on our manuscript. We address each major comment below and outline revisions that will strengthen the empirical support for our claims regarding practical fairness.

read point-by-point responses

  1. Referee: [§3 (HEFA formulation)] §3 (HEFA formulation): The hierarchical KL terms are motivated as correcting inter- and intra-group imbalances that a global objective leaves unaddressed. However, the construction depends on the chosen group partition; if groups are defined at a coarse level (e.g., top-level categories), the intra-group term only equalizes within those buckets. The paper must show that the learned additive adjustments do not create or leave uncorrected exposure monopolization at finer partitions (e.g., by popularity strata or item metadata) that are not part of the hierarchy, as this directly affects whether the method delivers the claimed practical fairness.

    Authors: We agree that HEFA's effectiveness is tied to the practitioner-specified group hierarchy, which is a deliberate design choice to enable flexible adaptation to diverse fairness requirements. The method does not claim to automatically correct imbalances at arbitrary finer partitions outside the chosen hierarchy. To directly address the concern, the revised manuscript will include an additional analysis of exposure distributions and monopolization metrics on finer partitions (popularity strata and item metadata) not used in training. This will confirm that the learned adjustments do not introduce new imbalances at those levels. revision: yes

  2. Referee: [§4 (Experiments)] §4 (Experiments): The headline result of 'substantial fairness gains with negligible accuracy loss' is supported by aggregate metrics. To substantiate that HEFA plus the joint NDCG term does not trade off against finer imbalances, the results must include (a) exposure variance or max-min ratios within each defined group and (b) an ablation or secondary evaluation on sub-partitions not used in the hierarchical loss. Without these, the claim that the method 'enables flexible adaptation to diverse fairness requirements' remains incompletely supported.

    Authors: We concur that aggregate metrics alone leave room for stronger validation of finer-grained balance. The revised experiments section will report within-group exposure variance and max-min ratios for each defined group on all three datasets. We will also add an ablation evaluating the method on sub-partitions (e.g., popularity-based strata) excluded from the hierarchical loss. These additions will provide direct evidence that the joint NDCG optimization preserves balance at finer levels, better supporting the flexibility claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation or claims

full rationale

The paper introduces PFA as a post-hoc adapter that learns additive score adjustments from embeddings and optimizes them via a composite loss (global KL on exposure plus HEFA hierarchical inter/intra-group terms plus differentiable NDCG). Fairness gains and accuracy preservation are asserted on the basis of experiments on three public datasets, not by algebraic reduction of the objective to itself. No equations are presented in the abstract that equate a derived quantity to a fitted input by construction, no load-bearing self-citations are invoked to establish uniqueness or force the architecture, and the method is motivated by explicit limitations of prior retraining or reranking approaches. The derivation chain therefore remains self-contained and externally falsifiable via the reported empirical evaluation.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on two domain assumptions: that user-item embeddings from the frozen model contain enough signal to learn useful fairness adjustments, and that a target fair exposure distribution can be meaningfully specified for any desired fairness requirement. No free parameters or invented entities are named in the abstract.

axioms (2)
  • domain assumption User-item embeddings from the frozen recommender contain sufficient information to learn effective additive fairness adjustments.
    The adapter is defined to learn from these embeddings.
  • domain assumption A target fair exposure distribution can be defined that is appropriate for diverse fairness requirements.
    The KL-divergence objective is defined with respect to such a target.

pith-pipeline@v0.9.0 · 5569 in / 1397 out tokens · 60293 ms · 2026-05-09T17:57:19.356533+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

65 extracted references · 65 canonical work pages

  1. [1]

    Himan Abdollahpouri and Robin Burke. 2019. Multi-stakeholder Recommenda- tion and its Connection to Multi-sided Fairness. InRMSE@RecSys, Vol. 2440

    work page 2019

  2. [2]

    Kenneth E. Batcher. 1968. Sorting Networks and Their Applications. InAFIPS, Vol. 32. 307–314

    work page 1968

  3. [3]

    Chi, and Cristos Goodrow

    Alex Beutel, Jilin Chen, Tulsee Doshi, Hai Qian, Li Wei, Yi Wu, Lukasz Heldt, Zhe Zhao, Lichan Hong, Ed H. Chi, and Cristos Goodrow. 2019. Fairness in Recommendation Ranking through Pairwise Comparisons. InKDD. 2212–2220

    work page 2019

  4. [4]

    Biega, Krishna P

    Asia J. Biega, Krishna P. Gummadi, and Gerhard Weikum. 2018. Equity of Attention: Amortizing Individual Fairness in Rankings. InSIGIR. 405–414

    work page 2018

  5. [5]

    Mathieu Blondel, Olivier Teboul, Quentin Berthet, and Josip Djolonga. 2020. Fast Differentiable Sorting and Ranking. InICML, Vol. 119. 950–959

    work page 2020

  6. [6]

    Ludovico Boratto, Francesco Fabbri, Gianni Fenu, Mirko Marras, and Giacomo Medda. 2024. Robustness in Fairness Against Edge-Level Perturbations in GNN- Based Recommendation. InECIR. 38–55

    work page 2024

  7. [7]

    Ludovico Boratto, Gianni Fenu, and Mirko Marras. 2021. Interplay between upsampling and regularization for provider fairness in recommender systems. User Model. User Adapt. Interact.31, 3 (2021), 421–455

    work page 2021

  8. [8]

    Miaomiao Cai, Lei Chen, Yifan Wang, Haoyue Bai, Peijie Sun, Le Wu, Min Zhang, and Meng Wang. 2024. Popularity-Aware Alignment and Contrast for Mitigating Popularity Bias. InKDD. 187–198

    work page 2024

  9. [9]

    Jiawei Chen, Hande Dong, Xiang Wang, Fuli Feng, Meng Wang, and Xiangnan He

    work page

  10. [10]

    ACM Trans

    Bias and Debias in Recommender System: A Survey and Future Directions. ACM Trans. Inf. Syst.41, 3 (2023), 67:1–67:39

    work page 2023

  11. [11]

    Xiao Chen, Wenqi Fan, Jingfan Chen, Haochen Liu, Zitao Liu, Zhaoxiang Zhang, and Qing Li. 2023. Fairly Adaptive Negative Sampling for Recommendations. In WWW. 3723–3733

    work page 2023

  12. [12]

    Sunhao Dai, Chen Xu, Shicheng Xu, Zhongxiang Sun, Liang Pang, Zhenhua Dong, and Jun Xu. 2025. Trustworthy Information Retrieval in the LLM Era: Bias, Unfairness, and Hallucination. InSIGIR-AP. 442–446

    work page 2025

  13. [13]

    Yashar Deldjoo and Tommaso Di Noia. 2025. CFaiRLLM: Consumer Fairness Evaluation in Large-Language Model Recommender System.ACM Trans. Intell. Syst. Technol.16, 6 (2025), 142:1–142:31

    work page 2025

  14. [14]

    Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, Jing Yi, Weilin Zhao, Xiaozhi Wang, Zhiyuan Liu, Hai-Tao Zheng, Jianfei Chen, Yang Liu, Jie Tang, Juanzi Li, and Maosong Sun. 2023. Parameter-efficient fine-tuning of large-scale pre-trained language models.Nat. Mach. Intell.5, 3 (20...

    work page 2023

  15. [15]

    Itay Eilat and Nir Rosenfeld. 2023. Performative Recommendation: Diversifying Content via Strategic Incentives. InICML, Vol. 202. 9082–9103

    work page 2023

  16. [16]

    Yingqiang Ge, Shuchang Liu, Ruoyuan Gao, Yikun Xian, Yunqi Li, Xiangyu Zhao, Changhua Pei, Fei Sun, Junfeng Ge, Wenwu Ou, and Yongfeng Zhang. 2021. Towards Long-term Fairness in Recommendation. InWSDM. 445–453

    work page 2021

  17. [17]

    Yingqiang Ge, Juntao Tan, Yan Zhu, Yinglong Xia, Jiebo Luo, Shuchang Liu, Zuohui Fu, Shijie Geng, Zelong Li, and Yongfeng Zhang. 2022. Explainable Fairness in Recommendation. InSIGIR. 681–691

    work page 2022

  18. [18]

    Elizabeth Gómez, Carlos Shui Zhang, Ludovico Boratto, Maria Salamó, and Mirko Marras. 2021. The Winner Takes it All: Geographic Imbalance and Provider (Un)fairness in Educational Recommender Systems. InSIGIR. 1808–1812

    work page 2021

  19. [19]

    Aditya Grover, Eric Wang, Aaron Zweig, and Stefano Ermon. 2019. Stochastic Optimization of Sorting Networks via Continuous Relaxations. InICLR

    work page 2019

  20. [20]

    Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. 2024. Parameter-Efficient Fine-Tuning for Large Models: A Comprehensive Survey. Trans. Mach. Learn. Res.2024 (2024)

    work page 2024

  21. [21]

    Emrul Hasan, Mizanur Rahman, Chen Ding, Jimmy Xiangji Huang, and Shaina Raza. 2026. Review-based Recommender Systems: A Survey of Approaches, Challenges and Future Perspectives.ACM Comput. Surv.58, 1 (2026), 25:1–25:41

    work page 2026

  22. [22]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yong-Dong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. InSIGIR. 639–648

    work page 2020

  23. [23]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochastic Opti- mization. InICLR

    work page 2015

  24. [24]

    Bell, and Chris Volinsky

    Yehuda Koren, Robert M. Bell, and Chris Volinsky. 2009. Matrix Factorization Techniques for Recommender Systems.Computer42, 8 (2009), 30–37

    work page 2009

  25. [25]

    Yunqi Li, Hanxiong Chen, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2021. User-oriented Fairness in Recommendation. InWWW. 624–632

    work page 2021

  26. [26]

    Zheda Mai, Ping Zhang, Cheng-Hao Tu, Hong-You Chen, Quang-Huy Nguyen, Li Zhang, and Wei-Lun Chao. 2025. Lessons and Insights from a Unifying Study of Parameter-Efficient Fine-Tuning (PEFT) in Visual Recognition. InCVPR. 14845– 14857

    work page 2025

  27. [27]

    Masoud Mansoury, Himan Abdollahpouri, Mykola Pechenizkiy, Bamshad Mobasher, and Robin Burke. 2022. A Graph-Based Approach for Mitigating Multi-Sided Exposure Bias in Recommender Systems.ACM Trans. Inf. Syst.40, 2 (2022), 32:1–32:31

    work page 2022

  28. [28]

    Marco Morik, Ashudeep Singh, Jessica Hong, and Thorsten Joachims. 2020. Con- trolling Fairness and Bias in Dynamic Learning-to-Rank. InSIGIR. 429–438

    work page 2020

  29. [29]

    Rahmani, and Yashar Deldjoo

    Mohammadmehdi Naghiaei, Hossein A. Rahmani, and Yashar Deldjoo. 2022. CP- Fair: Personalized Consumer and Producer Fairness Re-ranking for Recommender Systems. InSIGIR. 770–779

    work page 2022

  30. [30]

    Patro, Arpita Biswas, Niloy Ganguly, Krishna P

    Gourab K. Patro, Arpita Biswas, Niloy Ganguly, Krishna P. Gummadi, and Abhi- jnan Chakraborty. 2020. FairRec: Two-Sided Fairness for Personalized Recom- mendations in Two-Sided Platforms. InWWW. 1194–1204. Post-hoc Provider Fairness Adaptation via Hierarchical Exposure Alignment SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia

    work page 2020

  31. [31]

    Felix Petersen, Christian Borgelt, Hilde Kuehne, and Oliver Deussen. 2021. Dif- ferentiable Sorting Networks for Scalable Sorting and Ranking Supervision. In ICML, Vol. 139. 8546–8555

    work page 2021

  32. [32]

    Felix Petersen, Christian Borgelt, Hilde Kuehne, and Oliver Deussen. 2022. Mono- tonic Differentiable Sorting Networks. InICLR

    work page 2022

  33. [33]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme

    work page

  34. [34]

    BPR: Bayesian Personalized Ranking from Implicit Feedback. InUAI. 452– 461

    work page

  35. [35]

    Xiaoyu Shi, Quanliang Liu, Hong Xie, Yanan Bai, and Mingsheng Shang. 2024. Maximum Entropy Policy for Long-Term Fairness in Interactive Recommender Systems.IEEE Trans. Serv. Comput.17, 3 (2024), 1029–1043

    work page 2024

  36. [36]

    Smith, Aishwarya Satwani, Robin Burke, and Casey Fiesler

    Jessie J. Smith, Aishwarya Satwani, Robin Burke, and Casey Fiesler. 2024. Rec- ommend Me? Designing Fairness Metrics with Providers. InFAccT. 2389–2399

    work page 2024

  37. [37]

    Yifan Wang, Weizhi Ma, Min Zhang, Yiqun Liu, and Shaoping Ma. 2023. A Survey on the Fairness of Recommender Systems.ACM Trans. Inf. Syst.41, 3 (2023), 52:1–52:43

    work page 2023

  38. [38]

    Yibiao Wei, Jie Zou, Weikang Guo, Guoqing Wang, Xing Xu, and Yang Yang

    work page

  39. [39]

    MSCRS: Multi-modal Semantic Graph Prompt Learning Framework for Conversational Recommender Systems. InSIGIR. 42–52

    work page

  40. [40]

    Yibiao Wei, Jie Zou, Pengfei Zhang, Xiao Ao, Weikang Guo, Zeyu Ma, and Yang Yang. 2026. CARD: Non-Uniform Quantization of Visual Semantic Unit for Generative Recommendation. InSIGIR. 1–12

    work page 2026

  41. [41]

    Hongyi Wen, Xinyang Yi, Tiansheng Yao, Jiaxi Tang, Lichan Hong, and Ed H. Chi. 2022. Distributionally-robust Recommendations for Improving Worst-case User Experience. InWWW. 3606–3610

    work page 2022

  42. [42]

    Guoli Wu, Zhiyong Feng, Shizhan Chen, Hongyue Wu, Xiao Xue, Jianmao Xiao, Guodong Fan, Hongqi Chen, and Jingyu Li. 2025. FairSort: Learning to Fair Rank for Personalized Recommendations in Two-Sided Platforms.IEEE Trans. Knowl. Data Eng.37, 2 (2025), 641–654

    work page 2025

  43. [43]

    Haolun Wu, Chen Ma, Bhaskar Mitra, Fernando Diaz, and Xue Liu. 2023. A Multi-Objective Optimization Framework for Multi-Stakeholder Fairness-Aware Recommendation.ACM Trans. Inf. Syst.41, 2 (2023), 47:1–47:29

    work page 2023

  44. [44]

    Haolun Wu, Bhaskar Mitra, Chen Ma, Fernando Diaz, and Xue Liu. 2022. Joint Multisided Exposure Fairness for Recommendation. InSIGIR. 703–714

    work page 2022

  45. [45]

    Jiancan Wu, Xiang Wang, Fuli Feng, Xiangnan He, Liang Chen, Jianxun Lian, and Xing Xie. 2021. Self-supervised Graph Learning for Recommendation. In SIGIR. 726–735

    work page 2021

  46. [46]

    Yao Wu, Jian Cao, Guandong Xu, and Yudong Tan. 2021. TFROM: A Two-sided Fairness-Aware Recommendation Model for Both Customers and Providers. In SIGIR. 1013–1022

    work page 2021

  47. [47]

    Chongjun Xia, Xiaoyu Shi, Hong Xie, Yun Lu, Pu Li, and Mingsheng Shang. 2026. Beyond Trade-offs: Leveraging Spatiotemporal Heterogeneity of User Preference for Long-term Fairness and Accuracy in Interactive Recommendation.ACM Trans. Web20, 1 (2026), 3:1–3:26

    work page 2026

  48. [48]

    Zhichen Xiang, Hongke Zhao, Chuang Zhao, Ming He, and Jianping Fan. 2024. Performative Debias with Fair-exposure Optimization Driven by Strategic Agents in Recommender Systems. InKDD. 3507–3517

    work page 2024

  49. [49]

    Chen Xu, Sirui Chen, Jun Xu, Weiran Shen, Xiao Zhang, Gang Wang, and Zhenhua Dong. 2023. P-MMF: Provider Max-min Fairness Re-ranking in Recommender System. InWWW. 3701–3711

    work page 2023

  50. [50]

    Chen Xu, Zhirui Deng, Clara Rus, Xiaopeng Ye, Yuanna Liu, Jun Xu, Zhicheng Dou, Ji-Rong Wen, and Maarten de Rijke. 2025. FairDiverse: A Comprehensive Toolkit for Fairness- and Diversity-aware Information Retrieval. InSIGIR. 3540– 3550

    work page 2025

  51. [51]

    Chen Xu, Yuxin Li, Wenjie Wang, Liang Pang, Jun Xu, and Tat-Seng Chua. 2025. Bridging Jensen Gap for Max-Min Group Fairness Optimization in Recommenda- tion. InICLR

    work page 2025

  52. [52]

    Chen Xu, Xiaopeng Ye, Wenjie Wang, Liang Pang, Jun Xu, and Tat-Seng Chua

    work page

  53. [53]

    A Taxation Perspective for Fair Re-ranking. InSIGIR. 1494–1503

    work page

  54. [54]

    Chen Xu, Xiaopeng Ye, Jun Xu, Xiao Zhang, Weiran Shen, and Ji-Rong Wen. 2025. LTP-MMF: Toward Long-Term Provider Max-Min Fairness under Recommenda- tion Feedback Loops.ACM Trans. Inf. Syst.43, 1 (2025), 11:1–11:29

    work page 2025

  55. [55]

    Chen Xu, Jujia Zhao, Wenjie Wang, Liang Pang, Jun Xu, Tat-Seng Chua, and Maarten de Rijke. 2025. Understanding Accuracy-Fairness Trade-offs in Re- ranking through Elasticity in Economics. InSIGIR. 539–548

    work page 2025

  56. [56]

    Lanling Xu, Zihan Lin, Jinpeng Wang, Sheng Chen, Wayne Xin Zhao, and Ji-Rong Wen. 2024. Promoting Two-sided Fairness with Adaptive Weights for Providers and Customers in Recommendation. InRecSys. 918–923

    work page 2024

  57. [57]

    Tao Yang, Zhichao Xu, and Qingyao Ai. 2023. Vertical Allocation-based Fair Exposure Amortizing in Ranking. InSIGIR-AP. 234–244

    work page 2023

  58. [58]

    Tao Yang, Zhichao Xu, Zhenduo Wang, and Qingyao Ai. 2023. FARA: Future- aware Ranking Algorithm for Fairness Optimization. InCIKM. 2906–2916

    work page 2023

  59. [59]

    Sirui Yao and Bert Huang. 2017. Beyond Parity: Fairness Objectives for Collabo- rative Filtering. InNeurIPS. 2921–2930

    work page 2017

  60. [60]

    Hyunsik Yoo, Zhichen Zeng, Jian Kang, Ruizhong Qiu, David Zhou, Zhining Liu, Fei Wang, Charlie Xu, Eunice Chan, and Hanghang Tong. 2024. Ensuring User-side Fairness in Dynamic Recommender Systems. InWWW. 3667–3678

    work page 2024

  61. [61]

    Meike Zehlike and Carlos Castillo. 2020. Reducing Disparate Exposure in Ranking: A Learning To Rank Approach. InWWW. 2849–2855

    work page 2020

  62. [62]

    Jizhi Zhang, Keqin Bao, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Is ChatGPT Fair for Recommendation? Evaluating Fairness in Large Language Model Recommendation. InRecSys. 993–999

    work page 2023

  63. [63]

    Kaike Zhang, Qi Cao, Fei Sun, Yunfan Wu, Shuchang Tao, Huawei Shen, and Xueqi Cheng. 2026. Robust Recommender System: A Survey and Future Directions. ACM Comput. Surv.58, 1 (2026), 10:1–10:38

    work page 2026

  64. [64]

    Aggarwal, and Tyler Derr

    Yuying Zhao, Yu Wang, Yunchao Liu, Xueqi Cheng, Charu C. Aggarwal, and Tyler Derr. 2025. Fairness and Diversity in Recommender Systems: A Survey. ACM Trans. Intell. Syst. Technol.16, 1 (2025), 2:1–2:28

    work page 2025

  65. [65]

    Jiacong Zhou, Xianyun Wang, and Jun Yu. 2024. Optimizing Preference Alignment with Differentiable NDCG Ranking.arXiv preprint arXiv:2410.18127(2024)

    work page arXiv 2024