Fairness in Recommender Systems: Research Landscape and Future Directions

• Recommender systems influence consumer choices, platform profitability, and the information that can shape individual, community, and societal beliefs, decisions, and actions.Source: §1 · p. 2
• Negative effects such as provider-profit optimization and misinformation spread have challenged the historical view of recommenders as benevolent value-creating systems [Jannach and Jugovac, 2019].Source: §1 · p. 2
• Fairness is a central ethical question in responsible recommendation, alongside broader debates about fair machine learning and fair AI [Ntoutsi et al., 2020; Trattner et al., 2022; Burke, 2017; Ekstrand et al., 2022; Barocas et al., 2019; Mehrabi et al., 2021].Source: §1 · p. 2
• Because fairness is socially constructed and subjective, competing stakeholder interests make a general definition of fair recommendation difficult or even impossible [Ekstrand et al., 2022; Abdollahpouri et al., 2020a; Naghiaei et al., 2022].Source: §1 · p. 2
• The survey analyzes more than 150 recent computer-science papers to examine fairness dimensions and definitions, application scenarios, and operationalization through methodologies, algorithms, and metrics.Source: §1 · pp. 2–3
• Research in computing typically assumes a clear fairness definition and frames fairness-aware recommendation as algorithm design to optimize a metric, motivating more faceted and multidisciplinary approaches.Source: §1 · p. 3

2.1 Examples of Unfair Recommendations

• Recommendations can unfairly prioritize platform profit, reinforce discrimination, disadvantage less-popular providers, or promote politically skewed or misleading content across stakeholders.Examples span e-commerce, employment, music streaming, news, and social media, involving consumers, artists, labels, platforms, and society [Abdollahpouri et al., 2020a].Source: §2.1 · pp. 3–4

2.2 Reasons for Unfairness

• Unfairness may originate in biased data, model choices, intentional objectives, or historical societal conditions, requiring interventions during preprocessing, learning, or post-processing.Models may reinforce popularity or protected-attribute biases, while fairness interventions require operationalizable goals such as target distributions or evaluation metrics.Source: §2.2 · pp. 5–6

2.3 Notions of Fairness

• Fairness has multiple, sometimes incompatible notions—including individual versus group, process versus outcome, direct versus indirect, statistical versus predictive, static versus dynamic, and associative versus causal.The literature review finds that most works do not clarify the normative claim being addressed, its claimant, or relevant public debate [Cooper, 2020].Source: §2.3 · pp. 6–8

2.4 Related Concepts: Responsible Recommendation and Biases

• Fairness is a dimension of responsible recommendation, while bias is not inherently unfair because its ethical significance depends on underlying normative claims.Popularity bias may overexpose certain items without necessarily constituting ethical or legal unfairness; related concerns include robustness, privacy, interpretability, accountability, transparency, safety, and ethics [Ekstrand et al., 2022].Source: §2.4 · p. 10

3.1 Paper Collection Process

• The survey used DBLP title searches, explicit inclusion and exclusion criteria, snowballing, and researcher expertise to identify relevant papers in an emerging field.Search terms covered fair recommendation, collaborative filtering, bias, ranking, and reranking.Source: §3.1 · p. 11
• After manually screening 268 results, the process retained 157 papers explicitly addressing fairness in recommender systems, excluding preprints, surveys, and fair information-retrieval studies.Each retained paper was read by at least two researchers and categorized across multiple dimensions.Source: §3.1 · p. 11

3.2 Relation to Previous Surveys

• Unlike surveys focused on taxonomies, technical approaches, or general fair AI, this study maps the research landscape, emphasizes evaluation and methodology, and quantifies gaps [2022; 2022b].Its analyses find a strong predominance of data-based experiments using abstract computational metrics without humans in the loop.Source: §3.2 · pp. 12–13

4.1 Publication Activity per Year

• Fairness research in recommender systems has grown consistently, with the earliest explicitly fairness-focused surveyed paper published in 2017 and 157 papers included overall.Figure 1 reports the number of surveyed papers per year.Source: §4.1 · p. 13, Figure 1 · p. 14

4.2 Types of Contributions

• Technical papers dominate the literature, especially algorithmic work on debiasing data and fairer recommendation through list re-ranking.The survey distinguishes technical and conceptual papers, with non-technical work including guidelines, user-perception studies, and auditing discussions.Source: §4.2 · p. 14
• Algorithmic fairness solutions commonly assume predefined mathematical goals, whereas practical challenges involve agreeing on fairness objectives and designing metrics that reliably represent them.Fairness goals may depend on environmental factors, stakeholder impacts, ethical questions, and the deployment context.Source: §4.2 · pp. 14–15

4.3 Categorization of Notions of Fairness in Literature

• Group fairness is more common than individual fairness, and only rarely are both considered; group analyses typically compare exposure or relevance across consumer, provider, or multi-stakeholder groups.Groups may be defined using external demographics or interaction-derived attributes, whose reliability is generally not analyzed.Source: §4.3 · pp. 16, 19
• Most surveyed studies address single-sided fairness, leaving multisided fairness in multistakeholder recommender systems as an important research gap.Consumer and provider fairness can conflict, although some work reports that both sides can be improved without compromising recommendation quality.Source: §4.3 · pp. 17, 21
• Static evaluation predominates, with only 16 works studying dynamic environments, while just three works investigate recommendation fairness using causal considerations.Dynamic studies use simulation or reinforcement learning to capture feedback and long-term effects; causal studies use counterfactual reasoning or causal graphs.Source: §4.3 · pp. 22–23, Figure 5 · p. 22, Figure 6 · p. 24

4.4 Application Domains and Datasets

• Video, music, e-commerce, and finance dominate evaluations, while MovieLens appears in more than one third of papers despite often lacking sensitive attributes.Dataset availability strongly shapes the field, making MovieLens a widely adopted and potentially overused case.Source: §4.4 · p. 24
• Studies operationalize fairness through item popularity, user preference calibration, recommendation accuracy groups, and domain-specific attributes such as price, gender, or business category.Examples include increasing long-tail exposure, matching users’ genre-diversity preferences, and grouping users by achieved recommendation accuracy.Source: §4.4 · p. 25
• Many settings remain artificial because suitable datasets from consequential domains are scarce, risking abstraction traps that omit complex interactions between recommender systems and real-world sociotechnical contexts.The authors argue that fairness claims must state their assumptions, hypotheses, and normative basis, since lower accuracy or less popular recommendations are not inherently unfair.Source: §4.4 · pp. 25–26

4.5 Methodology

• Among 83 technical papers, offline experiments predominate, whereas only 6 involved humans and 3 reported field tests, leaving user fairness perceptions comparatively understudied.The review identifies only one interview-based study and few papers combining experiment types; examples include Htun et al. [2021], Geyik et al. [2019], and Beutel et al. [2019].Source: §4.5 · p. 27, Figure 8 · p. 27
• Most technical papers propose fairness-enhancing algorithms, typically through modeling or post-processing, while preprocessing is comparatively rare and re-ranking is common.The review distinguishes data analysis or outcome comparison, algorithmic approaches, and new metrics or evaluation approaches; fairness steps occur before, within, or after ranking.Source: §4.5 · pp. 28–29, Figure 9 · p. 28, Figure 10 · p. 29
• The field lacks standardized methodology: researchers call for more analytical, domain-specific work because target distributions and metric values are rarely validated against stakeholder judgments.In job recommendations, Wang et al. [2022a] found that a debiasing algorithm improved fairness without reducing accuracy, yet participants preferred the original recommendations.Source: §4.5 · pp. 28–29
• Offline studies use varied group- and individual-fairness metrics centered on exposure, effectiveness, disparity, loss, deviation cost, and worst-off utility, reflecting differing stakeholder objectives and trade-offs.Exposure primarily concerns provider visibility, effectiveness concerns preference-aligned utility, and metrics may balance customer fairness, provider fairness, and overall accuracy.Source: §4.5 · pp. 29–30, Table 3 · p. 30, Table 4 · p. 33
• Human studies are rare, reproducibility remains limited because most surveyed works share neither code nor data, and causal, counterfactual, long-term, and dynamic fairness research is comparatively uncommon.Sü̈hr et al. [2021] observed mixed effects of fair rankings in job recommendation, while the survey identifies more research into causes of unfairness as a gap.Source: §4.5 · pp. 35, 37, §4.6 · p. 39
• The methodology review finds that fairness metrics may oversimplify application problems, conflate bias with unfairness, and omit normative claims about desirable exposure, parity, calibration, or performance trade-offs.These concerns include uncertainty about which long-tail items deserve exposure, how target distributions are justified, and whether accuracy disparities are acceptable in realistic use cases.Source: §4.5 · p. 36

4.6 Landscape Overview

• The taxonomy organizes fairness research by implementation, target representation, benefit, measurement, fairness level, affected stakeholders, time horizon, and causes of unfairness.The landscape is presented as a multi-faceted framework for describing fairness research in recommender systems.Source: §4.6 · p. 37, Figure 12 · p. 40
• In-processing and post-processing methods, typically re-ranking, dominate because they are easier to apply to existing recommendation systems, while evaluation is mainly quantitative through offline experiments or A/B tests.Qualitative approaches are reported as almost never used.Source: §4.6 · p. 38
• Fairness metrics commonly target catalog size, relevance, or parity, with parity-based representations most common even though the appropriate target depends on the application domain.Fairness benefits may emphasize exposure or relevance, with relevance additionally requiring effective alignment with user preferences.Source: §4.6 · p. 38
• Group-level fairness is prevalent because it is easier to operationalize than individual fairness, while fairness research balances consumer and provider perspectives but largely overlooks intersectionality.One reported exception is Shen et al. [2023], which applies gender and skin-color intersectionality to conversational recommendation.Source: §4.6 · p. 39

Main observations

• Fairness research in recommender systems predominantly develops fairness-aware algorithms, while user interfaces, human-in-the-loop methods, and social-science perspectives remain limited.The authors argue that recommender systems should support human values rather than optimize abstract computational fairness metrics alone.Source: §5 · p. 40
• The field addresses individual and group fairness and extends fairness considerations to items and providers, whose treatment ultimately affects people and organizations.This broadens the scope beyond traditional fair-ML applications that primarily focus on people.Source: §5 · pp. 40–41

Research limitations

• Media datasets, especially MovieLens, dominate technical evaluations, revealing a shortage of real-world datasets and frequent reliance on synthetic or artificially protected groups.The available datasets constrain which fairness aspects can reasonably be studied.Source: §5 · p. 41
• Offline experiments are the method of choice, while few studies involve humans; operationalization often oversimplifies fairness by equating popularity biases with unfairness and assuming a gold-standard target exposure distribution.Some works re-brand beyond-accuracy research as fairness research while missing clear normative claims, increasing the danger of abstraction traps with limited practical impact.Source: §5 · p. 41

Future directions

• Future research should develop realistic fairness definitions, appropriate data and counterfactual evaluation, reciprocal-setting models, and practical fairness auditing.Key challenges include incompatible fairness notions and user perceptions, missing sensitive attributes addressed through augmentation or simulation [Rastegarpanah et al., 2019; Deldjoo et al., 2021b; Geyik et al., 2019], stakeholder equilibrium in reciprocal systems, and auditing constraints requiring stronger policies [Koshiyama et al., 2022; Krafft et al., 2020].Source: §5 · pp. 41–43
• Because current metrics may not represent fairness in particular contexts, the field needs multidisciplinary research on definitions, stakeholder perceptions, and real-world problem settings.The authors specifically call for collaboration across computer science, psychology, social science, and subject-matter expertise, paralleling concerns in explainable AI.Source: §5 · pp. 43–44