Fairness in Recommendation: A Survey

Yunqi Li, Hanxiong Chen, Shuyuan Xu, Yingqiang Ge, Juntao Tan, Shuchang Liu, Yongfeng ZhangarXiv 2022 · OpenReview

Introduction

Recommender systems (RS) serve as the invisible engines of the modern digital economy, steering user choices in domains ranging from entertainment to high-stakes employment. However, as data-driven entities, these systems are prone to inheriting and amplifying societal biases—such as gender or racial discrimination—or introducing new algorithmic biases like popularity bias.

Despite an explosion in research, the field of "fairness" remains fragmented across information retrieval, data mining, and ethics. This survey provides a unified framework to bridge these gaps, offering a systematic map of why unfairness occurs, how it is defined, and how it can be mitigated.

A Systematic Taxonomy

To organize the diverse landscape of fairness research, the authors propose a comprehensive taxonomy that categorizes interventions and definitions. This structure allows researchers to navigate the complex interplay between the causes of bias and the technical methods used to address them.

Taxonomy of Fairness in Machine Learning

Methods of Intervention

Fairness interventions are typically integrated into the standard machine learning pipeline at three distinct stages:

  • Pre-processing: This involves modifying training data before model training. Techniques include re-weighting samples or mapping data to a transformed space to neutralize inherent biases.
  • In-processing: Fairness is embedded directly into the model's objective function. Common approaches include Regularization (penalizing unfairness), Adversarial Learning (using a discriminator to ensure sensitive attributes are not learned), and Reinforcement Learning (optimizing for long-term fairness in dynamic environments).
  • Post-processing: This "black-box" approach re-ranks the output of a trained model to satisfy specific fairness constraints, such as ensuring proportional representation of protected groups in a Top-K list.

A significant advancement highlighted in the survey is the use of Causal Methods, which utilize causal graphs and counterfactual reasoning to move beyond simple statistical correlations and address the root causes of unfairness.

Key Insights: Philosophy and Complexity

The survey emphasizes that technical metrics are often grounded in classical theories of justice. For instance, "Equal Opportunity" in machine learning serves as a technical implementation of egalitarian justice.

However, achieving fairness is complicated by several factors:

  • Orthogonality: Group fairness (treating different groups similarly) and individual fairness (treating similar individuals similarly) can be mutually exclusive.
  • Multi-Stakeholder Dynamics: RS must balance the needs of users (quality of service), items/producers (exposure), and the platform (revenue).
  • Static vs. Dynamic Effects: Static fairness constraints can sometimes backfire in the long run as user preferences and item popularities evolve in a closed-loop system.

The primary contribution of this survey is the synthesis of a rapidly expanding field. The data shows a clear upward trajectory in research activity, with over 260 works categorized.

Growth of research papers in fairness for recommendation systems

The authors also highlight the "Pareto-optimal" trade-off inherent in most methods: increasing fairness often results in a slight decrease in recommendation accuracy (e.g., NDCG or Hit Rate). Interestingly, some recent studies suggest that prioritizing fairness can improve long-term user retention, suggesting a potential alignment between ethics and business goals.

Future Directions and Limitations

Despite recent progress, several challenges remain:

  • Large Language Models (LLMs): The integration of LLMs into recommendation introduces subtle biases where sensitive attributes are leaked through indirect linguistic cues.
  • Lack of Consensus: There is no universal definition of fairness, necessitating research into methods that can satisfy multiple conflicting requirements.
  • Real-World Evaluation: Most research is conducted offline. There is a critical need for live A/B testing and better simulation environments (e.g., RecSim) to understand the long-term impact of fairness interventions in production settings.

Summary updated

BibTeX
@misc{li2022fairness,
  author = {Yunqi Li and Hanxiong Chen and Shuyuan Xu and Yingqiang Ge and Juntao Tan and Shuchang Liu and Yongfeng Zhang},
  title = {{Fairness in Recommendation: A Survey}},
  year = {2022},
  url = {https://openreview.net/forum?id=UD1-By75Un},
}
Figure