[Paper Review] Fairness in Recommendation: Foundations, Methods and Applications

Yunqi Li, Hanxiong Chen+5 more|arXiv (Cornell University)|May 26, 2022

Explainable Artificial Intelligence (XAI)Computer Science18 citations

TL;DR

This survey provides a comprehensive, systematic overview of fairness in recommender systems, synthesizing foundational concepts from machine learning, classifying fairness definitions, reviewing key techniques, and curating datasets. It identifies core challenges such as data sparsity, multi-stakeholder fairness, and privacy-utility trade-offs, while advocating for unified evaluation protocols and simulation environments to advance fair recommendation research.

ABSTRACT

As one of the most pervasive applications of machine learning, recommender systems are playing an important role on assisting human decision making. The satisfaction of users and the interests of platforms are closely related to the quality of the generated recommendation results. However, as a highly data-driven system, recommender system could be affected by data or algorithmic bias and thus generate unfair results, which could weaken the reliance of the systems. As a result, it is crucial to address the potential unfairness problems in recommendation settings. Recently, there has been growing attention on fairness considerations in recommender systems with more and more literature on approaches to promote fairness in recommendation. However, the studies are rather fragmented and lack a systematic organization, thus making it difficult to penetrate for new researchers to the domain. This motivates us to provide a systematic survey of existing works on fairness in recommendation. This survey focuses on the foundations for fairness in recommendation literature. It first presents a brief introduction about fairness in basic machine learning tasks such as classification and ranking in order to provide a general overview of fairness research, as well as introduce the more complex situations and challenges that need to be considered when studying fairness in recommender systems. After that, the survey will introduce fairness in recommendation with a focus on the taxonomies of current fairness definitions, the typical techniques for improving fairness, as well as the datasets for fairness studies in recommendation. The survey also talks about the challenges and opportunities in fairness research with the hope of promoting the fair recommendation research area and beyond.

Motivation & Objective

• To address the fragmented state of fairness research in recommender systems by providing a unified, systematic overview.
• To establish foundational knowledge on fairness in machine learning, particularly in classification and ranking, as a basis for understanding fairness in recommendation.
• To classify and organize existing fairness definitions in recommender systems into a coherent taxonomy for clearer conceptual understanding.
• To review and categorize state-of-the-art techniques for improving fairness in recommendation, including both algorithmic and data-level approaches.
• To identify critical open challenges—such as privacy-preserving data collection, lack of diverse benchmark datasets, and the need for unified evaluation frameworks—while proposing future research directions.

Proposed method

• Surveying and synthesizing over 100 works on fairness in recommendation to establish a structured overview of the field.
• Introducing a taxonomy of fairness definitions in recommender systems, distinguishing between user-level, item-level, and group-level fairness concerns.
• Categorizing fairness techniques into three main classes: pre-processing (e.g., data reweighting), in-processing (e.g., adversarial training, fairness regularizers), and post-processing (e.g., re-ranking).
• Curating and presenting publicly available datasets with sensitive attributes (e.g., gender, race, income) to support future fairness research.
• Proposing the development of simulation environments—inspired by RecSim—for evaluating dynamic and long-term fairness effects in recommendation systems.
• Advocating for privacy-preserving data curation methods such as federated learning and differential privacy to enable fair learning without compromising user privacy.

Experimental results

Research questions

• RQ1 What are the core fairness definitions and taxonomies used in recommender systems, and how do they differ from fairness in classification and ranking?
• RQ2 What are the main technical approaches (pre-processing, in-processing, post-processing) for achieving fairness in recommendation, and how do they compare in effectiveness?
• RQ3 What are the key challenges in collecting and using sensitive user and item attributes for fairness research, especially concerning privacy and data availability?
• RQ4 How can unified evaluation protocols be developed to fairly compare different fairness methods across multiple fairness criteria?
• RQ5 What role can simulation platforms play in evaluating long-term and dynamic fairness in recommendation systems?

Key findings

• Fairness in recommender systems is inherently more complex than in standard machine learning tasks due to multiple stakeholders, dynamic environments, and data sparsity.
• Existing fairness definitions in recommendation span user, item, and group fairness, with significant variation in formulation and evaluation, making cross-method comparison difficult.
• The current landscape lacks sufficient publicly available datasets with sensitive attributes, limiting reproducibility and scalability of fairness research.
• There is a critical tension between privacy protection and fairness promotion, as sensitive user data are often required but also highly vulnerable to misuse.
• A unified evaluation framework that simultaneously measures fairness, accuracy, explainability, robustness, and privacy is still missing and represents a key research gap.
• Simulation platforms such as RecSim are promising for dynamic fairness evaluation, and future work should extend such frameworks to support fairness-specific benchmarking.