Fairness in Recommendation: Foundations, Methods and Applications

1 INTRODUCTION

• Unfair recommendations can discriminate against user groups, concentrate exposure on popular sellers, or over-represent particular political ideologies.These outcomes can harm multiple stakeholders and undermine a positive, sustainable recommendation ecosystem.Source: §1 · pp. 1–2
• The survey uses classification and ranking as background because recommendation problems can often be formulated as either task.It introduces fairness concepts and methods in those tasks before examining recommendation-specific concerns.Source: §1 · p. 2
• Recommendation fairness has unique challenges, including multiple models and goals, long-term effects, sparse data, and multiple stakeholders.The survey addresses these complexities through a taxonomy, techniques, datasets, and open challenges.Source: §1 · p. 2

2 RELATED SURVEYS

• Existing fair-ranking surveys often focus on item exposure, whereas recommender fairness can involve multiple sides.The survey also distinguishes its scope from surveys on recommendation bias that primarily target accuracy or robustness rather than fairness.Source: §2 · p. 3
• Earlier surveys cover fairness in classification, ranking, or particular machine-learning areas, while this survey focuses on recommendation fairness.The paper notes that two recommendation fairness surveys were released concurrently [63].Source: §2 · p. 3
• The survey combines technical and philosophical foundations with a comprehensive taxonomy of fairness notions in recommendation.It also covers task-specific techniques and datasets suited to recommendation fairness research.Source: §2 · p. 3

3 FAIRNESS IN MACHINE LEARNING

• Fairness in machine learning is introduced through causes of unfairness, fairness definitions, and fair-learning methods.The section also uses a mind map to show the basic structure of fairness in machine learning and recommender systems.Source: §3 · p. 4

3.1 The Causes of Unfairness

• Unfairness can stem from data or model bias, and conflicting fairness requirements can also explain why one requirement is violated.Biases can interact through a feedback loop between user behavior, model outputs, and subsequent training data.Source: §3.1 · p. 4, §3.1.3 · p. 5, §3.1.2 · p. 5

3.2 Fairness Definition

• Fairness definitions include group, individual, and hybrid forms, and the choice depends on the scenario and fairness requirements.Group fairness compares treatment across groups, individual fairness concerns similar individuals, and hybrid fairness seeks multiple requirements that may conflict.Source: §3.2 · p. 5, §3.2.1 · p. 6, §3.2.2 · p. 6

3.3 Methods for Fair Machine Learning

• Fair-learning methods fall into preprocessing, in-processing, and post-processing, each with distinct trade-offs and constraints.Preprocessing changes data, in-processing modifies learning, and post-processing transforms model outputs; their limitations include possible accuracy loss, optimization challenges, or test-time access to sensitive features.Source: §3.3 · p. 6, §3.3.1 · p. 7, §3.3.2 · p. 7

4 PHILOSOPHICAL FOUNDATIONS OF FAIRNESS IN MACHINE LEARNING: THE THEORIES OF JUSTICE

• The survey relates fairness research in machine learning and recommendation to the philosophical concept of justice.It follows prior discussions of justice as a philosophical origin of machine-learning fairness and extends the discussion to recommendation [211].Source: §4 · pp. 7–8

4.1 The Theories of Justice

• Theories of justice include utilitarianism, contractarianism, and egalitarianism, with many machine-learning fairness notions aligned with egalitarianism [211].Egalitarianism emphasizes substantively equal distribution, while machine-learning fairness must specify which individuals or groups count as equal and what treatment is appropriate.Source: §4.1 · p. 8

4.2 The Distinctions of Justice

• Justice distinctions connect fairness to ideals, resource distribution, decision procedures, and comparisons between people or groups.The survey contrasts conservative with ideal, corrective with distributive, procedural with substantive, and comparative with non-comparative justice.Source: §4.2 · pp. 8–10

4.3 The Scope of Justice

• The scope of justice concerns who is covered and when justice principles apply, while machine-learning fairness generally focuses on human interests within specific domains.Item-side recommendation fairness concerns human interests such as producer profits.Source: §4.3 · p. 10

5.1 Fairness Concerns

• Group fairness compares protected and advantaged groups using predicted-positive-rate parity or confusion-matrix rates; examples include Statistical Parity and Equal Opportunity.The survey also describes Equalized Odds, which requires equal true- and false-positive rates across groups.Source: §5.1.1 · p. 11
• Individual fairness asks that similar people receive similar predictions, using either a causal counterfactual criterion or similarity measured over non-sensitive features.Counterfactual Fairness compares predictions in real and counterfactual worlds; Fairness Through Awareness bounds prediction differences by individuals’ distance.Source: §5.1.2 · p. 12

5.2 Methods

• Fair classification methods fall into pre-processing, in-processing, and post-processing, each suited to different access, model-modification, and deployment requirements.Pre-processing changes data, in-processing adds fairness objectives during training, and post-processing modifies scores or outputs after training.Source: §5.2 · p. 12, §5.2.3 · pp. 13–14
• Pre-processing alters labels, transforms data, or perturbs features; massaging and re-weighting reduce sensitive-feature dependence while potentially sacrificing accuracy.Massaging changes labels across sensitive groups while preserving the overall positive-class probability; re-weighting changes subjects’ influence.Source: §5.2.1 · p. 12
• In-processing incorporates fairness constraints into training, while post-processing adjusts scores or feature weights and can be used with a fixed model.In-processing allows a flexible fairness–accuracy trade-off but requires model changes; post-processing is model-agnostic but limited to changing output probabilities.Source: §5.2.2 · p. 13, §5.2.3 · pp. 13–14
• Classification experiments on statistical-parity constraints found improved fairness but sacrificed accuracy, with a threshold controlling the accuracy–unfairness trade-off.The optimization is described as convex, with a Pareto-optimal trade-off between classifier accuracy and the fairness constraint.Source: §5.2.2 · p. 13

6 FAIRNESS IN RANKING

• Fair ranking informs recommendation, but ranking studies usually focus on candidate items whereas recommendation fairness spans multiple stakeholders [34].Source: §6 · p. 14

6.1 Fairness Concerns

• Probability-based fairness bounds the number or proportion of protected candidates in the top-K list and is designed for group fairness [14, 41, 93, 253].Source: §6.1.1 · p. 15
• Exposure/attention-based fairness addresses position bias by relating group exposure to relevance, and can formalize group or individual fairness.Individual unfairness can also be measured by comparing two candidates’ exposure [255].Source: §6.1.2 · p. 15

6.2 Methods

• A pre-processing approach maps individuals to prototype representations using cluster-assignment probabilities, then combines utility and fairness losses; on learning-to-rank it achieved best individual fairness but worse utility than baselines.Source: §6.2.1 · p. 16
• In-processing can impose cross-group pairwise equal opportunity while maximizing pairwise accuracy; experiments report reduced fairness violation at the cost of lower test AUC.The paper claims the algorithm can apply to any pairwise metric.Source: §6.2.2 · pp. 16–17
• Fairness improvements can come at a cost: methods reduce fairness violations or improve individual fairness while lowering test AUC, accuracy, or utility.The reported utility cost for the pre-processing method is specifically on learning-to-rank.Source: §6.2.1 · p. 16, §6.2.2 · p. 17, §6.2.3 · p. 17
• Post-processing learns a probabilistic ranking that maximizes utility under fairness constraints, with document exposure weighted by position importance; statistical parity can also lower accuracy.Source: §6.2.3 · p. 17

7.1 Preliminaries of Recommender Systems

• Recommenders learn user–item preferences from interaction histories and rank items for users; fairness concerns can involve user features such as gender or race and item features such as category or popularity.Source: §7.1 · p. 18

7.2 Examples of Unfairness in Recommendation

• In an e-commerce example, the top 5% most active users had higher F1 scores than the remaining 95%, while similarly relevant items received unequal exposure at different ranking positions.The item example uses the exposure drop-off 1/log(1 + i).Source: §7.2 · pp. 18–19

7.3 A Taxonomy of Fairness Notions in Recommendation

• The taxonomy distinguishes recommendation-specific notions, including user fairness, multi-sided fairness, and dynamic fairness, from notions such as group, individual, and causal fairness used more broadly across machine learning.Source: §7.3 · pp. 19–20
• Fairness can concern groups or individuals and users or items; multi-sided approaches seek to address multiple stakeholders, such as customers and producers, together.The survey describes approaches that combine minimum item exposure with user utility guarantees and others that balance exposure transitions with minimum customer utility.Source: §7.3.1 · p. 20, §7.3.2 · p. 21, §7.3.3 · pp. 21–22
• Most recommendation fairness work uses association-based measures, while causal fairness remains less studied and can be difficult to compute from observational data.Causal approaches use structural knowledge, interventions, or counterfactuals; the survey identifies causality as an opportunity for further recommendation research.Source: §7.3.4 · pp. 22–23
• Static fairness is common, but changing user preferences, item popularity, and group labels make dynamic and long-term fairness important; applying static criteria myopically can exacerbate unfairness.Source: §7.3.5 · pp. 23–24, §7.3.6 · p. 24
• The taxonomy also distinguishes population-wide from personalized fairness, explainable from blackbox fairness, and centralized from federated fairness.Explainable fairness is especially relevant given recommenders’ many explicit and latent features; federated fairness raises tensions between protecting sensitive user data and evaluating or optimizing user fairness, as well as collaboration fairness.Source: §7.3.7 · p. 24, §7.3.8 · p. 25, §7.3.9 · p. 26

7.4 Methods

• Fair-recommendation methods first specify fairness metrics and then select techniques suited to those metrics, which vary across fairness definitions and research settings.Source: §7.4 · p. 26
• In recommendation, fairness criteria can be formulated as regularizers or constraints to guide optimization; objectives may maximize utility under fairness constraints, fairness under utility bounds, or jointly optimize fairness and utility.Examples include a pairwise-comparison regularizer to improve ranking fairness and a group-recommendation problem formulated as multiple-objective optimization with fairness as a regularizer.Source: §7.4.1 · pp. 26–27
• Adversarial learning trains a recommendation predictor against a sensitive-feature classifier to remove sensitive information from representations, offering model-agnostic flexibility but often difficult, unstable training [13] [21] [71] [74] [238].Applications include private-attribute protection and decomposing user embeddings into bias-aware and bias-free components for fair news ranking.Source: §7.4.2 · pp. 27–28
• Reinforcement-learning methods address long-term, dynamic fairness by modeling recommendation as an MDP, but must guard against user tampering through manipulation of preferences and beliefs [148] [256].Examples include dynamically constraining item exposure, learning fairness-utility trade-offs, and representing user preferences and system fairness status in the MDP state.Source: §7.4.3 · p. 28
• Causal methods model relationships between sensitive variables and decisions to address discrimination, popularity or selection bias, and counterfactual fairness, while relying on assumptions or graphs that may be unavailable [109].They can also support decision transparency and counterfactual explanations of item-exposure unfairness.Source: §7.4.4 · pp. 28–29
• Other approaches use antidote-data augmentation, VAE sampling noise for long-term fairness with an acceptable recommendation-quality trade-off, and self-distillation to improve fairness for cold users [224].These methods respectively avoid modifying the original data or algorithm, mitigate position bias across recommendation rounds, and regularize predictions on behavior-augmented data.Source: §7.4.5 · p. 29

7.5 Datasets and Settings for Fairness in Recommendation Research

• Fairness datasets require features suited to the fairness definition, but datasets with user or item sensitive attributes are limited; examples include MovieLens and Last.FM with user attributes and Amazon with item categories.The survey collects publicly available datasets covering recommendation tasks and user or item attributes.Source: §7.5.1 · pp. 29–31
• Fair-recommendation evaluations vary by fairness goal and task, typically requiring both recommendation-accuracy and fairness metrics, with data preprocessing and feature engineering aligned to the chosen fairness definition.Accuracy measures include NDCG, Hit Rate, Precision, Recall, and F1-score; fairness measures include group-level equalized odds, equal opportunity, or statistical parity differences and individual-level treatment differences.Source: §7.5.2 · pp. 31–32

8.1 Lack of Consensus on Fairness Definitions

• Fairness lacks a consensus definition because both who counts as equal and what counts as equal treatment depend on context.Fairness demands and the biases that produce unfairness also vary across scenarios and perspectives.Source: §8.1 · p. 32
• Methods usually target one fairness requirement, but multiple demands may arise together and some fairness notions inherently conflict.The survey calls for studying unified approaches in limited cases and finding reasonable balances where requirements conflict.Source: §8.1 · pp. 32–33
• Personalized fairness could let users express their own demands, while scenario-specific evaluation could help select suitable fairness notions.The survey raises comparisons such as group versus individual fairness and static versus dynamic fairness.Source: §8.1 · p. 33
• Different fairness metrics across studies make methods difficult to compare, motivating unified evaluation metrics or benchmark datasets.Source: §8.1 · p. 33

8.2 Relationship between Fairness and other Trustworthy AI Perspectives

• When fairness demands conflict, explanations could clarify why some users receive unfair treatment now and how compensation may occur later.The survey also notes that explaining why a decision is fair to a user may support trust in the system.Source: §8.2 · p. 33
• The survey proposes studying explainable fairness and how explainability and fairness can benefit each other.Source: §8.2 · p. 33
• Other proposed connections include fairness with robustness and out-of-distribution generalization, privacy and federated learning, and controllability of users’ fairness demands.Source: §8.2 · p. 33

8.3 Causal Foundations for Fairness

• Causal fairness in recommendation remains limited, and classification techniques may not transfer directly even when recommendation is formulated as classification.Source: §8.3 · p. 33
• Collaborative models can encode relationships between user features and behavior through interaction data, despite not directly using features as inputs.Thus, omitting sensitive features is not a sufficient route to counterfactual fairness in these models; the survey calls for further causal methods.Source: §8.3 · p. 34

8.4 Understanding the Connection between Bias and Fairness

• Bias can contribute to unfair recommendation, but its effects depend on the source and form of bias, and bias does not necessarily produce unfairness.A single training-data bias may yield individual, group, short-term, or long-term unfairness; calibration may avoid unfairness despite data bias.Source: §8.4 · p. 34
• Debiasing research typically targets accuracy or robustness, whereas fairness work specifies fairness metrics and evaluates both fairness and accuracy.Fairness promotion may impose outcome constraints rather than use debiasing, with a risk of losing accuracy.Source: §8.4 · p. 34
• The survey calls for clarifying the bias–unfairness relationship and connecting debiasing with fairness-promotion methods to better understand causes and improve both fairness and accuracy.Source: §8.4 · p. 35

8.5 Fairness of Large Language Model based Recommendation

• Indirectly sensitive prompt words such as “pretty” or “young” may lead LLM-based recommendation models to make unfair recommendations.These words may indirectly reveal a user’s gender or age, even when the prompt omits explicit sensitive-feature terms.Source: §8.5 · p. 35
• P5 formulates recommendation tasks as language generation using prompts, enabling one model, loss, and data format across multiple tasks [91] [108].The tasks include rating prediction, sequential and direct recommendation, explanation generation, and review summarization.Source: §8.5 · p. 35
• UP5 uses Counterfactually-Fair-Prompting to develop an unbiased foundation model for fairness-aware recommendation [107].Other early work evaluates ChatGPT and reports that it still exhibits unfairness for some sensitive attributes when generating recommendations [258].Source: §8.5 · p. 35

8.6 Considerations in Real-World Deployment

• User-oriented deployment delivers fairness-aware results to platform users, including consumers, sellers, viewers, and creators, while developer-oriented deployment provides tools for developers and policy makers.The latter tools support understanding system unfairness without immediately changing results delivered to users.Source: §8.6 · p. 35
• LinkedIn Talent Search’s fairness framework improved fairness metrics without affecting business metrics in online A/B testing.It targeted equality of opportunity and demographic parity in candidate ranking.Source: §8.6 · pp. 35–36
• Amazon SageMaker Clarify provides deployed toolkits for developers to detect and monitor bias in data or models and assess whether outputs are fair.The survey calls for studying stakeholder fairness demands and how developer-oriented deployment knowledge can transfer to user-oriented deployment.Source: §8.6 · p. 36

8.7 Understanding the Relationship between Fairness and Utility

• Complex real-world systems make fairness problems difficult to identify, while companies may prioritize business metrics over fair treatment when the two conflict.The survey also notes legal requirements as one way to promote fair treatment and calls for further research on fairness and utility.Source: §8.7 · p. 36
• Future work should examine when fairness and utility can improve together, and how to minimize, explain, or distribute losses when they conflict.The proposed directions include multi-objective and Pareto optimization, explainable and dynamic fairness, and user interfaces.Source: §8.7 · p. 36
• Static datasets can show short-term fairness–accuracy trade-offs, but do not establish that increasing fairness harms utility in deployed systems over the long term.The survey calls for real-world long-term study, noting that fair treatment may support user retention, interest, trust, and engagement, and eventually platform utility and profits.Source: §8.7 · p. 37

8.8 Developing Better Metrics, Benchmarks and Simulation Environments

• Fairness research datasets remain limited because studies need relevant features, including sensitive personal information, whose use and publication require care.The survey points to anonymization and privacy-protection methods such as federated learning or differential privacy as possible approaches [17].Source: §8.8 · p. 37
• Users may withhold or misreport personal information because of unfair-treatment or privacy concerns, creating a dilemma for fairness work.The survey suggests building trust and incentives, improving incomplete data, and developing methods that do not rely on unavailable sensitive features.Source: §8.8 · p. 37
• Because fairness definitions have distinct evaluation methods, the survey proposes a common protocol for comparing fairness requirements and evaluating fairness alongside utility and other trustworthy considerations.These considerations include accuracy, explainability, robustness, and privacy.Source: §8.8 · p. 37
• Recent recommendation comparisons cover only user fairness, and the survey identifies fairness simulation platforms as a useful research resource [110].The survey notes the difficulty of measuring many fairness aspects within one framework.Source: §8.8 · p. 38

• The survey provides background on fairness in machine learning, including its causes, methods, and differing considerations.Source: §9 · p. 38
• It introduces classification and ranking fairness definitions and methods to clarify concepts related to recommender systems.Source: §9 · p. 38
• It classifies recommender-system fairness definitions in a taxonomy and surveys technical methods, evaluation methods, and datasets.Source: §9 · p. 38
• It discusses challenges and opportunities for fairness research in recommendation.Source: §9 · p. 38