Source-linked AI summary
Fairness in Recommender Systems: Research Landscape and Future Directions
Yashar Deldjoo, Dietmar Jannach, Alejandro Bellogin, Alessandro Difonzo, Dario Zanzonelli
TL;DR
Recommender systems can shape information exposure, decisions, and stakeholder value, while fairness research remains conceptually and operationally unsettled. This survey reviews the field’s concepts and research practices across more than 160 publications, finding that computer-science work often relies on abstract algorithmic fairness metrics while leaving normative claims underexplained. The authors therefore identify a need for broader, more interdisciplinary approaches, subject to the survey’s documented scope boundaries.
Problem
Fairness research in recommender systems lacks sufficiently developed treatment of what constitutes a fair recommendation in particular contexts and how competing stakeholder interests should be addressed.
Method
The survey reviews fundamental fairness concepts and analyzes more than 160 scholarly publications to map research methodologies, fairness measures, and algorithmic approaches.
Results
The literature predominantly uses fairness-aware algorithms and abstract computational fairness metrics, while comparatively limited work examines user interfaces, human-in-the-loop approaches, or disciplines beyond computer science.
Takeaways & Limitations
Fairness-aware recommendation research should address the underlying normative claims and application context through more faceted and multidisciplinary approaches.
Takeaways & Limitations
The survey’s algorithmic-fairness focus often leaves historical discrimination and the broader social reasons why characteristics are protected outside its main focus.
Abstract
from arXiv · showhide
Recommender systems can strongly influence which information we see online, e.g., on social media, and thus impact our beliefs, decisions, and actions. At the same time, these systems can create substantial business value for different stakeholders. Given the growing potential impact of such AI-based systems on individuals, organizations, and society, questions of fairness have gained increased attention in recent years. However, research on fairness in recommender systems is still a developing area. In this survey, we first review the fundamental concepts and notions of fairness that were put forward in the area in the recent past. Afterward, through a review of more than 160 scholarly publications, we present an overview of how research in this field is currently operationalized, e.g., in terms of general research methodology, fairness measures, and algorithmic approaches. Overall, our analysis of recent works points to certain research gaps. In particular, we find that in many research works in computer science, very abstract problem operationalizations are prevalent and questions of the underlying normative claims and what represents a fair recommendation in the context of a given application are often not discussed in depth. These observations call for more interdisciplinary research to address fairness in recommendation in a more comprehensive and impactful manner.
1 Introduction
Fairness in recommender systems matters because recommendations can influence stakeholders’ outcomes, beliefs, and actions. This survey reviews the emerging field and finds that computing research often treats fairness as abstract metric optimization.
- Recommendations can influence consumer choices, platform profitability, and the information shaping beliefs, decisions, and actions.
- Awareness has grown that recommenders may produce negative effects, including provider-oriented promotions and misinformation spread.
- Fairness is difficult to define generally because it is subjective and involves competing stakeholder interests.
- The survey analyzes more than 150 recent computer-science papers across fairness dimensions, application scenarios, methodologies, algorithms, and metrics.
- The survey identifies research gaps and discusses potential shortcomings and future directions.
- Computing research typically assumes a clear fairness definition and frames the problem as designing algorithms to optimize a metric.
2 Background and Foundations
Fairness in recommender systems involves multiple stakeholders, possible sources of unfairness, and competing conceptions of what fairness means. The surveyed literature often leaves normative claims underexplained and operationalizes fairness through pipeline interventions and technical metrics.
- Fairness concerns include provider profit maximization, misinformation, discrimination, and other effects across common recommendation settings.
- Recommendation services can affect consumers, providers, artists, labels, platforms, and society, whose goals may diverge.
- Unfairness may arise from society, data, models, evaluation, deployment, or user interaction, including biased sampling and popularity reinforcement.
- Fairness interventions can occur during data preprocessing, model learning and optimization, or post-processing.
- Fairness is a complex societal construct with numerous perspectives and no exhaustive definition.
- Most surveyed works do not clarify the normative claim being addressed or who represents or makes that claim.
- Researchers usually do not connect recommendation-fairness questions to specific public discussions outside computer science.
3 Research Methodology
The survey uses a mixed, semi-systematic process to identify and categorize fairness research in recommender systems. It intentionally provides a broad research landscape rather than a strict systematic review or a technical synthesis of specific methods.
- Paper identification: The authors identify papers through DBLP queries using predefined search terms and explicit inclusion and exclusion criteria, then supplement results with snowballing and researcher experience.The search terms were chosen to cover an emerging field whose terminology is not yet unified.
- Screening and inclusion: After manually screening 268 papers, the survey retains 157 papers that explicitly address fairness and recommender systems.Papers focused on fair information retrieval were excluded, and included papers were read by at least two researchers.
- Analysis: The survey categorizes the included papers across multiple dimensions to analyze how fairness research is operationalized.The categorization follows the paper-identification and screening process described for the survey corpus.
- Scope and positioning: The work is not intended as a strict systematic literature review, but as an outline of a broader picture of current research activities.The authors describe their approach as mixed and semi-systematic.
- Relation to previous surveys: Unlike related surveys that review or synthesize approaches, this study quantifies which fairness notions are studied and how the research problem is operationalized.This orientation is used to identify less frequently studied areas and potential research gaps.
4 Landscape of Fairness Research in Recommender Systems
The surveyed literature is dominated by technical, algorithm-focused work, with offline evaluations and media datasets especially prevalent. Across fairness dimensions, group and single-sided evaluations are more common than individual, multisided, dynamic, or causal analyses, revealing several research gaps.
- Research scope and contributions: 157 papers were included in the survey, and publication activity on fairness in recommender systems has grown in recent years.The reviewed papers were categorized across multiple dimensions of fairness, methodology, applications, and datasets.
- Types of contributions: Technical papers dominate the literature, especially algorithmic approaches such as debiasing and fair list re-ranking.The survey contrasts these works with conceptual contributions, including design guidelines, user-perception studies, and auditing discussions.
- Fairness notions: Group fairness is more frequently studied than individual fairness, and only rarely are both considered together.Group fairness commonly compares privileged and unprivileged groups across exposure or relevance and consumer or provider stakeholders.
- Fairness notions: Single-sided fairness dominates multisided fairness research, despite recommendation settings often affecting multiple stakeholders.The survey identifies multisided fairness in multistakeholder applications as an important research gap.
- Fairness notions: Static evaluations are predominant: only 16 surveyed works consider dynamically changing environments, even though feedback loops and delayed effects can matter.Longitudinal analysis therefore requires alternatives such as synthetic-data experiments or simulation-based methodologies.
- Fairness notions: Only three surveyed works investigate recommendation fairness through causality, while calibration fairness addresses deviations between user interests and recommendations.Causal approaches include counterfactual explanations, causal analysis of protected attributes, and causal graphs for visual bias.
5 Discussion
The survey finds that fairness research in recommender systems has broadened across fairness notions and affected stakeholders, but remains dominated by abstract algorithmic and offline evaluations. It identifies gaps involving realistic definitions, human-in-the-loop research, interdisciplinary perspectives, and longitudinal evaluation.
- Research orientation: Most published works are algorithmic, with technical contributions predominantly consisting of fairness-aware algorithms.
- Research orientation: Researchers address individual and group fairness and have expanded fairness considerations to items and providers.
- Methodological gaps: Media datasets, including MovieLens, dominate technical evaluations, although their fairness dimensions appear limited.
- Methodological gaps: Offline dataset experiments are the method of choice, while only very few works involve human-in-the-loop studies.
- Open challenges: Static offline evaluations do not capture longitudinal dynamics when fairness goals or algorithm outputs change over time.
- Open challenges: Fairness definitions can conflict, making it difficult to balance competing fairness notions with recommendation performance and user perception.