PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on the Fairness of Recommender Systems

Yifan Wang, Weizhi Ma, Min Zhang, Yiqun Liu, Shaoping Ma

arXiv:2206.03761v2cs.IR

TL;DR

Fairness research in recommender systems is scattered across complex resource-allocation processes and varied definitions, despite the systems’ societal role. This survey systematically reviews the field, finding that group fairness is the most common target and that consistent and calibrated fairness are prominent definitions. It also identifies limited investigation of emerging scenarios such as short-video recommendation.

  • Problem

    Fairness in recommendation is difficult to synthesize because systems involve multiple complex resource-allocation processes and varied fairness definitions.

  • Method

    The survey systematically reviews formally published fairness research and organizes definitions, measurements, datasets, methods, and future directions.

  • Results

    Group fairness is the most common target, while consistent fairness and calibrated fairness are prominent definitions in existing recommendation studies.

  • Takeaways & Limitations

    The survey provides a structured overview of fairness definitions, evaluation resources, methods, and directions for future recommendation research.

  • Takeaways & Limitations

    Existing studies lack investigation of some emerging recommendation scenarios, including short-video recommendation.

Abstract

from arXiv · show

Recommender systems are an essential tool to relieve the information overload challenge and play an important role in people's daily lives. Since recommendations involve allocations of social resources (e.g., job recommendation), an important issue is whether recommendations are fair. Unfair recommendations are not only unethical but also harm the long-term interests of the recommender system itself. As a result, fairness issues in recommender systems have recently attracted increasing attention. However, due to multiple complex resource allocation processes and various fairness definitions, the research on fairness in recommendation is scattered. To fill this gap, we review over 60 papers published in top conferences/journals, including TOIS, SIGIR, and WWW. First, we summarize fairness definitions in the recommendation and provide several views to classify fairness issues. Then, we review recommendation datasets and measurements in fairness studies and provide an elaborate taxonomy of fairness methods in the recommendation. Finally, we conclude this survey by outlining some promising future directions.

1 INTRODUCTION

Recommendation systems help address information overload while allocating exposure and information, making fairness important for users, items, and platforms. This survey organizes scattered fairness research by definitions, measurements, datasets, methods, and future directions.

  • Recommendation systems alleviate information overload but also allocate information to users and exposure to items.
  • Studies report fairness differences across users’ recommendation accuracy, diversity, and novelty, with female and older users receiving worse results in some settings.
  • Minority items can receive worse ranking performance and less exposure, while paid items may receive worse platform services than non-paid items.
  • Unfairness can harm users, niche providers, content diversity, and platforms’ long-term interests, making its mitigation a critical recommender-system issue.
  • Fairness in recommendation is distinctive because systems serve users and items, operate through feedback loops, personalize outputs, and jointly optimize accuracy with other measures.
  • The survey reviews more than 60 fairness-related recommendation papers and synthesizes definitions, measurements, datasets, methods, and future research directions.

2 DEFINITIONS OF FAIRNESS IN RECOMMENDATION

The survey organizes fairness in recommendation through multiple classification views, distinguishing process from outcome fairness and grouping outcomes by target and fairness concept. It finds that existing work mainly studies outcome fairness at the group level, especially consistent and calibrated fairness.

  • Scope: The survey presents several fairness definitions because no consensus exists and the literature spans diverse interpretations.It introduces definitions appearing in recommendation research and illustrates them in a taxonomy.
  • Primary distinctions: Fairness is classified by whether attention centers on the allocation process or its outcome.Process fairness concerns the model and information used, whereas outcome fairness concerns the fairness of received information or exposure.
  • Target: Outcome fairness is further divided into group fairness and individual fairness according to the target of evaluation.Group fairness compares outcomes across groups, while individual fairness evaluates outcomes at the individual level.
  • Target: Group fairness may require attention to subgroups formed by multiple attributes, including cases where single-attribute groups appear fair.The survey also notes that group memberships can be dynamic and that individual fairness can be viewed as a special case of group fairness.
  • Fairness concepts: Fairness concepts specify what requirements make outcomes fair, including consistent, calibrated, envy-free, and counterfactual fairness.Consistent fairness treats similar individuals or groups similarly, while calibrated fairness makes outcome value proportional to merit; the other concepts impose distinct outcome constraints.
  • Research focus: Existing research emphasizes outcome fairness, mainly group fairness, with consistent and calibrated fairness the dominant concepts; a few studies examine Rawlsian maximin fairness.The survey reports this distribution in its lookup table of reviewed fairness definitions.

3 VIEWS OF FAIRNESS IN RECOMMENDATION

The survey classifies recommendation fairness across allocation processes using views such as fairness subjects, granularity, and optimization objects. It distinguishes user, item, and joint fairness, as well as single versus amortized and treatment-based versus impact-based fairness.

  • The survey organizes recommendation fairness by subjects, granularity, and optimization objects across multiple allocation processes.These views and corresponding reviewed work are summarized in Table 4.
  • Subject: The survey reports that previous work mainly studies item and user fairness, with relatively little work addressing joint fairness.
  • Subject: User fairness concerns comparable treatment across users, whereas item fairness concerns comparable treatment or relevance-proportional allocation across items.The survey notes that provider fairness can be treated as group-level item fairness.
  • Subject: Joint fairness considers users and items together, but improving item fairness can worsen or leave user fairness unchanged.The survey characterizes this conflict as making joint fairness challenging.
  • Granularity: Fairness granularity distinguishes single-list requirements from amortized requirements aggregated across multiple recommendation lists.Prior studies concentrate on amortized fairness because single fairness may be unachievable in some scenarios.
  • Optimization Object: Treatment-based fairness evaluates allocated treatments such as predicted scores or exposure, while impact-based fairness evaluates recommendation effects such as user feedback.The survey reports that most previous studies focus on treatment-based fairness because impact-based fairness cannot directly control user feedback.

4 MEASUREMENTS OF UNFAIRNESS IN RECOMMENDATION

The survey reviews fairness measurements by fairness definition, target, granularity, and application scope. It emphasizes that commonly used metrics are broadly applicable but often coarse-grained, while specialized metrics address particular fairness issues.

  • Fairness metrics differ in scope of application because fairness definitions, subjects, granularity, and optimization objects differ.The survey organizes reviewed measurements according to these conditions in Table 5.
  • Most fairness metrics target outcome fairness, reflecting its prominence in existing recommendation fairness research.The survey therefore presents more detailed metrics for consistent and calibrated fairness separately.
  • Common metrics apply broadly to users and items but are relatively coarse-grained and often summarize groups using only first-order information.They also fail to account for distinct characteristics of user fairness and item fairness.
  • Specialized metrics can better fit particular fairness issues, including statistical tests for population distributions and aggregation across items or ranking positions.
  • Consistent Fairness: KS statistics address higher-order inconsistency by comparing empirical utility distributions, with lower values indicating fairer recommendations.Utilities are often defined as predicted ratings for groups.
  • Item Fairness: Exposure metrics such as rND, rKL, and rRD incorporate ranking position through normalized discounted cumulative unfairness calculations.The survey reports that rKD is smoother and more robust than rRD.

4.3 Metrics for Calibrated Fairness (CA)

Calibrated fairness measures compare recommendation utility with a predefined merit or target distribution. The survey reviews divergence, distance, disparity, skew, and norm-based metrics across group and individual settings.

  • Calibrated fairness defines the merit of an individual or group and evaluates allocation against the resulting fair distribution.The merit function determines the fair allocation proportion for individuals or groups.
  • Group Fairness: KL-divergence compares utility and target distributions, such as item-group exposure and historical exposure, with lower values indicating fairer recommendations.
  • Group Fairness: NDKL computes KL-divergence at each ranking position and aggregates the values using normalized discounted cumulative calculation.Lower NDKL values indicate fairer recommendations.
  • Group Fairness: JS-divergence offers a symmetrical alternative to asymmetrical KL-divergence for comparing probability distributions.The survey notes that some work uses JS-divergence instead of KL-divergence.
  • Group Fairness: Overall disparity measures average disparity between utility proportions and merit proportions across groups using exposure-based or click-based utility.Lower values indicate fairer recommendations.
  • Individual Fairness: L1-norm treats merit and utility distributions as vectors and measures their distance, commonly for individual-level calibrated fairness.It is also used in some group-level studies, with lower values indicating fairer recommendations.
  • Some calibrated and consistent fairness measures are interconvertible by setting the target distribution or utility function appropriately.

4.4 Metrics for Other Fairness Definitions

The survey covers measurements for envy-free, Rawlsian maximin, maximin-shared, and process fairness. These measures evaluate preferences, worst-case utilities, satisfaction thresholds, or representation independence in different recommendation settings.

  • Envy-Free Fairness: Envy-free fairness compares users’ utilities for recommendation lists or items, with lower average envy or higher envy-free proportions indicating fairer outcomes.Definitions differ between group recommendations and general recommendations.
  • Envy-Free Fairness: In group recommendations, m-envy-free fairness requires a user to be envy-free for at least m items in the recommendation package.It is impossible for all users to be envy-free for all items in a group package.
  • Envy-Free Fairness: Envy measures based on predicted preferences may diverge from true preferences, making true-preference fairness difficult to measure offline.True-preference envy-free fairness requires counterfactual questions; bandit-based auditing is proposed as a complement.
  • Maximin Fairness: Rawlsian maximin fairness focuses on the worst individual or group and can use average utility among the bottom n% to improve robustness to noise.Higher values indicate fairer recommendations.
  • Maximin Fairness: Maximin-shared fairness measures the proportion of individuals whose outcomes exceed their maximin share, with higher proportions indicating fairer recommendations.For item fairness, the maximin share is defined as average exposure.
  • Process Fairness: Process fairness evaluates whether allocation inputs or representations are independent of fairness-related attributes.A classifier can predict such attributes from representations, with classification measurements used as fairness measures.

5.1 Overview of Fairness Methods

The survey organizes fairness methods by their position in the recommendation pipeline: data-oriented, ranking, and re-ranking methods. Different method types address overlapping fairness issues, although process and counterfactual fairness are addressed only by adversarial learning in the reviewed work.

  • Method taxonomy: Fairness methods are categorized into data-oriented, ranking, and re-ranking methods according to their position in the recommendation pipeline.Data-oriented methods change training data, ranking methods modify models or learning objectives, and re-ranking methods adjust model outputs.
  • Method distribution: Only a few data-oriented methods are reviewed, while regularization and adversarial learning dominate ranking methods and slot-wise methods dominate re-ranking.Reinforcement learning has gained attention for modeling dynamics and long-term effects.
  • Fairness coverage: Each method type can address several fairness issues, and most fairness issues can be addressed by multiple method types.The survey summarizes these method–issue relationships in Table 8.
  • Fairness coverage: Process and counterfactual fairness are addressed only by adversarial learning, whereas Rawlsian maximin and maximin-shared fairness tend to use global-wise re-ranking.The survey attributes the latter pattern partly to the smaller amount of related work.
  • Open direction: The survey identifies designing additional methods for more specific fairness issues as an area worth exploring.This observation follows the reviewed concentration of certain issues in particular method types.

5.2 Data-oriented Methods

Data-oriented methods improve recommendation fairness by modifying training data rather than the original recommendation model. They offer low coupling and broad candidate-set access, but later pipeline stages can degrade their effects and some approaches are time-consuming.

  • Approach: Data-oriented methods modify training data to improve recommendation fairness.Examples include re-sampling user groups and adding optimized antidote data during training.
  • Examples: Re-sampling on Movielens 1M alleviated unfairness, but not significantly.The method adjusts the proportion of different user groups in the training data.
  • Examples: Antidote-data optimization can better mitigate unfairness than re-sampling, but it is relatively more time-consuming.The antidote data can be updated by optimizing a fairness objective with gradient descent.
  • Advantages: These methods have low coupling because they do not require modifying the original recommendation model, and they face fewer candidate-set constraints.Operating at the front of the pipeline gives them potential to improve final recommendation fairness significantly.
  • Limitations: Their fairness performance may degrade through subsequent stages such as diversity re-ranking, making effective data-oriented design challenging.Multiple stages intervene between training data and final presentation.

5.3 Ranking Methods

Ranking methods modify recommendation models or optimization targets to learn fair representations or predictions. Regularization and adversarial learning are central approaches, while reinforcement learning supports dynamic and long-term fairness objectives, each with distinct limitations.

  • Overview: Ranking methods modify recommendation models or optimization targets to learn fair representations or recommendations.The survey divides them into regularization-based, adversarial learning-based, reinforcement learning-based, and other methods.
  • Reinforcement Learning: Reinforcement learning has gained attention because it can model dynamics and long-term effects, while some methods seek a Pareto frontier between fairness and utility.This supports controlling the fairness–utility trade-off for decision-makers.
  • Comparison and limitations: Ranking outputs are closer to final presentation than data-oriented outputs, but downstream re-ranking can still damage their fairness performance.The survey describes ranking-stage fairness improvements as more straightforward than data-stage improvements.
  • Regularization: Regularization adds fairness-related terms to recommendation loss functions, using either direct fairness metrics or indirect objectives.The general formulation is L = L_rec + λ·L_fair.
  • Regularization: Direct regularization is simple and effective but requires differentiable metrics, limiting application mainly to rating-prediction unfairness and often reducing recommendation performance.Exposure and ranking metrics are described as non-differentiable in this context.
  • Regularization: Indirect regularization can achieve better fairness and recommendation performance than direct methods, with FATR outperforming direct metric regularization on real datasets.FATR induces orthogonality between representations and fairness-related attributes through a regular term.
  • Adversarial Learning: Adversarial learning is dominant for fair representation learning, but its minimax optimization objective makes optimization difficult.Discriminators predict fairness-related attributes while the recommendation model learns representations that suppress that information.

5.4 Re-ranking Methods

Re-ranking methods adjust recommendation outputs after model scoring and are divided into slot-wise, user-wise, and global-wise approaches. They offer direct control over presented lists, but candidate-set limits and short-horizon decisions constrain some methods.

  • Taxonomy: Re-ranking methods adjust recommendation-model outputs and are categorized as slot-wise, user-wise, or global-wise.The three types differ in whether they add items one slot at a time, optimize for users directly, or re-rank users simultaneously.
  • Advantages: Re-ranking results are nearly identical to final presentation, making their fairness improvements straightforward, and they do not require changing recommendation models.These methods therefore have low coupling with the recommender system.
  • Limitations: A small candidate set can hamper re-ranking performance, and re-ranking cannot resolve fair-representation issues in the ranking stage.These limitations arise because re-ranking operates after candidate generation and ranking.
  • User-wise: User-wise and related approaches jointly optimize fairness and utility, often using a hyperparameter to control the trade-off.Some methods define objectives as linear combinations of recommendation performance and fairness utility.
  • User-wise: Personalized fairness scores can incorporate multiple item attributes and achieve a better fairness–utility trade-off, while single-attribute methods have narrower applicability.User tolerance differs across item attributes, motivating the multi-attribute formulation.
  • Global-wise and dynamic ranking: Dynamic re-ranking can theoretically guarantee fairness when the number of rankings is sufficiently large.A proportional-controller approach combines recommendation performance and fairness linearly.
  • Slot-wise: Slot-wise methods independently add items for each user and are intuitive and efficient, but their greedy decisions can be shortsighted and suboptimal.They optimize each position sequentially rather than jointly considering the full recommendation process.
  • Global-wise: Global-wise re-ranking is more suitable for user fairness because it re-ranks different users simultaneously.The reviewed results also report better fairness for one global-wise algorithm than an earlier algorithm.

6 DATASETS FOR FAIRNESS RECOMMENDATION STUDY

The survey finds that fairness-recommendation datasets are diverse in scenarios, interactions, scale, and available attributes, but coverage remains uneven. It catalogs attribute-based and attribute-free datasets while identifying emerging scenarios and missing sensitive attributes as priorities.

  • Dataset organization: The authors organize prior datasets by fairness-related attributes and by whether such attributes are available, presenting the reviewed resources in Table 9.Table 9 is a lookup table for datasets used in existing fairness research and lists attributes considered in prior work.
  • Attribute availability: Attribute-free datasets support fairness studies through individual-level criteria such as Rawlsian maximin fairness or interaction-based group construction.Interaction information can include user activity and item popularity, while datasets may contain only IDs and aligned feedback.
  • Dataset coverage: Existing datasets span movie, e-commerce, and job recommendations, with large- and small-scale data, impressions, clicks, ratings, and some multimodal information.Examples include Amazon, Movielens, Flixter, Netflix, ModCloth, Xing, and Yelp.
  • Fairness-related attributes: Researchers divide groups using dataset-specific attributes such as gender, age, item categories, and publishing information, or generic interaction-derived attributes such as activity and popularity.Generic attributes can be used when sensitive attributes are unavailable, including for privacy reasons.
  • Attribute availability: Generic interaction-derived attributes are dynamic, so users or items can move between protected and unprotected groups over time.An item may be popular now but previously cold, changing its group assignment.
  • Gaps and future data: The survey reports relatively rich and diverse datasets but notes insufficient coverage of emerging scenarios and attributes such as race.Short-video recommendation introduces multiple modalities, while new data collection must also address privacy concerns.

7 FUTURE DIRECTIONS

The survey identifies unresolved fairness challenges involving conflicting definitions, weak evaluation standards, incomplete data, joint user–item objectives, and deployment constraints. It proposes future work on scenario-specific consensus, benchmarks, broader metrics, causal methods, missing data, and efficient real-system algorithms.

  • Definitions: Fairness definitions can conflict, and no existing work determines their priority; suitable definitions may vary by scenario.The survey suggests that scenario-level consensus would be helpful, while acknowledging that a universal definition may not exist.
  • Evaluation: Existing fairness studies use varied measurements and data-processing strategies, motivating standardized benchmarks covering preprocessing, hyper-parameter tuning, and evaluation metrics.Without effective benchmarks, comparisons may be unfair and results may be non-reproducible.
  • Datasets: Fairness research lacks sufficient investigation of emerging scenarios such as short-video recommendation, whose multiple modalities differ from traditional settings.The survey specifically identifies a shortage of fairness-related work on short-video datasets.
  • Joint fairness: Future methods should jointly address user and item fairness while managing their potential conflict and the trade-off with recommendation accuracy.The survey discusses multi-objective learning, Pareto optimization, and learnable re-ranking as possible directions.
  • Beyond accuracy: Fairness evaluation should extend beyond accuracy to diversity, unexpectedness, and serendipity because unfairness has also been observed in these measurements.These metrics are connected to user satisfaction in the surveyed discussion.
  • Causal inference: Causal fairness remains underexplored in recommendation, especially for constructing and using causal graphs when models include interrelated additional features.The survey identifies causal-graph construction and removal of unfair-factor influence as open problems.
  • Missing data: Missing or unreliable fairness attributes can make existing methods ineffective, motivating approaches that improve fairness from partial information while reducing sensitive-information leakage risk.Users may report confidential or false gender information, preventing reliable identification of unfair treatment.
  • Real systems: Real systems require fairness methods that account for recall, ranking, and re-ranking, multiple objectives, short response times, and performance-related revenue risks.Diversity re-ranking may increase user unfairness, and fairness applied before re-ranking may therefore be ineffective.

8 CONCLUSION

The survey synthesizes fairness research in recommendation across definitions, perspectives, measurements, datasets, and methods. It finds group, consistent, and calibrated fairness to be especially prominent, while highlighting benchmark development and broader fairness methods as future needs.

  • Survey scope: The survey systematically reviews fairness-related recommendation research across definitions, views, measurements, datasets, and methods.It is intended to help readers understand fairness issues and provide research inspiration.
  • Definitions: Group fairness is the most common target, while consistent fairness and calibrated fairness are the most common fairness concepts.The survey classifies prior definitions according to different targets and concepts.
  • Views and measurements: The survey classifies fairness issues through perspectives including fairness subjects, granularity, and optimization objects, and summarizes representative measurements and common metrics.These views complement the survey’s taxonomy of fairness definitions.
  • Methods: Representative fairness methods are reviewed across data-oriented, ranking, and re-ranking approaches, with ranking and re-ranking more common than data adjustment.The conclusion states that only a few studies adjust data to improve fairness.
  • Resources and future work: The survey catalogs fairness-related datasets to help researchers locate relevant resources and outlines future work on definitions, benchmarks, joint user–item fairness, and metrics beyond accuracy.It also identifies explaining unfairness as a promising direction.
Loading 2206.03761v2…