Recommender systems leverage user interactions to predict their preferences and deliver personalized recommendations. Recent years have seen a great increase in their widespread usage in online areas, such as social media, e-commerce and even job applications. However, due to how these systems collect and learn from data, they are vulnerable to various biases, such as popularity bias, which also raises the question of their fairness for both users and providers. Researchers in the area have tried addressing the issue with various debiasing and fairness intervention methods, but these are often studied in separate strands of research, and the trade-off between fairness and accuracy is rarely explicitly evaluated when it comes to debiasing methods.
In this project, we replicate three state-of-the-art debiasing methods and analyze their impact on the fairness and accuracy of recommender models, particularly the trade-off between the two and how it can be controlled using hyper-parameters. We find that while the impact heavily depends on the method and dataset used, in many cases significant improvements can be made to fairness with little to no decrease in accuracy, using the right configuration of hyper-parameters.

Fairness in recommender systems is an increasingly critical concern as these models mediate access to information, opportunities, and visibility. While collaborative filtering (CF) approaches have been extensively scrutinized for popularity bias and unfair exposure, the fairness properties of content-based recommendation (CBR) models remain underexplored. In this work, we present a comparative evaluation of CF and CBR models—introducing a modular, feature-fused content-based recommender (MultiFuseCB)—on MovieLens 1M and Amazon Beauty datasets. We systematically analyze how the selection and weighting of content features, as well as the choice of embedding models, affect both recommendation accuracy and fairness, using metrics such as item coverage and popularity bias. Our results show that CBR models, with appropriate feature engineering, can achieve competitive accuracy while substantially improving fairness relative to CF baselines. We further demonstrate that certain features (e.g., year, genre, plot) and embedding choices can be leveraged to promote more equitable item exposure. These findings provide actionable insights for designing fairer content-based recommenders and highlight the importance of feature selection and model tuning in achieving both accuracy and fairness.

As important tools for information filtering, recommendation systems have greatly improved the efficiency of users' access to information in daily life by providing personalized suggestions. However, as people's reliance on it grows, recent studies have gradually revealed their potential risks of social unfairness, such as gender discrimination that may result from job recommendations. The unfairness not only harms the interests of specific individuals or groups but also threatens the credibility and long-term sustainability of systems. Therefore, building fairness-aware recommendation systems that proactively identify and mitigate unfairness is crucial for achieving responsible recommendation services. This study focuses on systematically evaluating the effectiveness of current fairness intervention strategies. Specifically, pre-processing methods (including data relabeling and resampling) and post-processing methods (including re-ranking, calibration, and equity of attention) are selected and implemented on the two datasets MovieLens-1M and Lastfm-NL, then comprehensively evaluated in terms of two types of metrics: accuracy and fairness. The experimental results show that different methods are effective in improving different fairness targets, with varying degrees of accuracy loss or gain. This paper further explores the trade-offs between maintaining accuracy and improving fairness on intervention methods, and proposes future improvement directions for fairness-aware recommendation systems in light of the experimental results.