Algorithmic Bias in Recommender Systems

Abstract

Recommender systems have seen considerable adoption in recent years, driven by modern streaming services, social media, and novel LLM applications. However, recommender systems show clear statistical biases and can expose stakeholders to social biases. This phenomenon is caused by several factors, such as the data, which tends to be too dirty for the statistical methods used in recommendation; the pipeline, which might directly introduce bias into recommendations; and the evaluation, which is often unaware of the underlying biases in the task. As such, we first reviewed current literature on the topic and discovered several discrepancies. Firstly, fairness-aware solutions, which aim to reduce social bias, are often not tested on Missing-at-Random (MAR) data, which is cleaner than the traditional Missing-not-at-Random (MNAR) data used in recommendations. Further, we establish a connection between statistical bias and social bias, and identify the need for user-group-based studies. As such, we first tested whether fairness-aware solutions benefit from MAR data similarly to debiasing solutions, which aim to reduce statistical bias. Then we investigated the extent to which debiasing solutions can address fairness issues, before finally delving into more detail on the individual user-group performance of some of our configurations. We found that some fairness-aware algorithms benefit from MAR data, though this does not appear to be universal. We also observed a noticeable benefit from diversification enabled by debiasing solutions, and we identified interesting insights into how interventions impact users based on the share of popular items they interact with.