Bias and Debias in Recommender System: A Survey and Future Directions (2010.03240v2)

Abstract: While recent years have witnessed a rapid growth of research papers on recommender system (RS), most of the papers focus on inventing machine learning models to better fit user behavior data. However, user behavior data is observational rather than experimental. This makes various biases widely exist in the data, including but not limited to selection bias, position bias, exposure bias, and popularity bias. Blindly fitting the data without considering the inherent biases will result in many serious issues, e.g., the discrepancy between offline evaluation and online metrics, hurting user satisfaction and trust on the recommendation service, etc. To transform the large volume of research models into practical improvements, it is highly urgent to explore the impacts of the biases and perform debiasing when necessary. When reviewing the papers that consider biases in RS, we find that, to our surprise, the studies are rather fragmented and lack a systematic organization. The terminology ``bias'' is widely used in the literature, but its definition is usually vague and even inconsistent across papers. This motivates us to provide a systematic survey of existing work on RS biases. In this paper, we first summarize seven types of biases in recommendation, along with their definitions and characteristics. We then provide a taxonomy to position and organize the existing work on recommendation debiasing. Finally, we identify some open challenges and envision some future directions, with the hope of inspiring more research work on this important yet less investigated topic. The summary of debiasing methods reviewed in this survey can be found at \url{https://github.com/jiawei-chen/RecDebiasing}.

Bias and Debias in Recommender Systems: A Comprehensive Overview

The paper "Bias and Debias in Recommender System: A Survey and Future Directions" presents a rigorous survey of biases present in recommender systems (RS) and explores methods for mitigating these biases. The authors organize the existing literature into a systematic taxonomy of biases and debiasing approaches, drawing attention to fragmented studies that have not consistently defined or addressed bias. By categorizing seven types of biases and corresponding debiasing methods, the paper provides a structured framework for future research in this area.

Types of Biases in Recommender Systems

The paper identifies seven key types of biases:

1. Selection Bias: Arising from users' selective behavior in rating items, leading to missing-not-at-random data. Approaches to address this include propensity score weighting and joint generative modeling to estimate missing data mechanisms.
2. Exposure Bias: Occurs because users can only interact with items they are exposed to, making unobserved interactions ambiguous. Methods such as exposure-based modeling and propensity scoring are often employed.
3. Conformity Bias: Derived from users’ tendencies to follow public opinion. Solutions involve disentangling conformity from true user preferences using models that incorporate social influence.
4. Position Bias: Users tend to interact with items at the top of a recommendation list regardless of relevance. Click models and position-based propensity scores are typical corrective strategies.
5. Inductive Bias: Introduced intentionally to models to improve generalization and efficiency. This includes assumptions and objectives that guide the learning process.
6. Popularity Bias: A bias towards recommending popular items more frequently than warranted based on their initial visibility, leading to the exacerbation of the "rich get richer" effect. Regularization techniques and adversarial methods help mitigate this bias.
7. Unfairness: Manifested when the system discriminates against certain user groups based on intrinsic or acquired traits. Approaches to address unfairness include regularization, adversarial learning, and causal modeling.

Implications for Future Research

The paper underscores the importance of developing comprehensive solutions that can address multiple biases simultaneously. Current solutions are often tailored to specific biases and may not generalize well when biases intersect.

Future work should aim to:

• Enhance the Accuracy of Propensity Scores: The efficiency of IPS-based methods relies heavily on accurate propensities, an area ripe for further exploration and refinement.
• Design General Debiasing Frameworks: Solutions that effectively handle combinations of biases represent a significant gap in current research.
• Improve Evaluation Metrics: Evaluating RS in an unbiased manner is crucial, and the use of hybrid evaluators combining both biased and unbiased data sources should be explored.
• Leverage Auxiliary Information: Incorporating auxiliary information and knowledge graphs can enhance debiasing strategies by providing rich contextual data.
• Apply Causal Modeling: Causal inference offers robust frameworks for interpreting the complexities of bias and its effect on recommendation predictions, providing avenues for explanation and deeper insights.
• Consider Dynamic Bias: Recognizing that biases evolve over time means future solutions should accommodate this dynamism in their models.
• Balance Fairness and Accuracy: Developing techniques that enable controlled trade-offs between system accuracy and fairness is essential for practical deployment.

Conclusion

The paper provides a scholarly foundation for understanding biases in RS and presents a roadmap for addressing these biases. By compiling diverse approaches under a unified framework, it offers an invaluable resource for researchers seeking to enhance the fairness and efficacy of recommender systems. With AI's perpetually evolving landscape, it is imperative to continue advancing debiasing methods to build more equitable and accurate systems.