| dc.description.abstract | Recommender Systems (RS) filter an immense array of options to provide us with the most suitable and relevant items. However, certain items are recommended excessively while others receive minimal exposure from the algorithms. Given the profound impact of recommender systems on our lives and decisions, essential questions arise about fairness and the equitable allocation of benefits and resources among all stakeholders.
The "popularity bias" characterizes the phenomenon where RS tend to disproportionately recommend popular items, thereby augmenting their exposure compared to less popular items. Given the relevance of less popular items to many users, popularity bias causes equity concerns. On one hand, the bias triggers a "rich-get-richer" effect, having tangible consequences for less popular artists, including reduced financial compensation and media attention for these artists. Fairness issues for users arise when the bias is responsible for varying levels of content quality for different user groups. Niche users, for example, might find popular recommendations less satisfying, although such recommendations are well-suited for mainstream-oriented users. Therefore, their satisfaction is directly linked to the popularity bias.
To investigate possibilities to mitigate the adverse effects of popularity bias, we develop a recommendation algorithm (RankALS) and employ popularity bias mitigation techniques. One aims at artist fairness (FA*IR), while the second focuses on user fairness (Calibrated Popularity). Through an algorithmic evaluation of the algorithms on the state-of-the-art music dataset (LFM-2b), we observe that the user-centric algorithm performs comparably to the base algorithm based on performance metrics (e.g., accuracy, NDCG) and recommends songs aligned with users' historical listening preferences in terms of popularity. This highlights its high user fairness. However, it retains an over-representation of popular items. Conversely, the artist-focused algorithm increases exposure for underrepresented songs, achieving item fairness, albeit at the expense of matching users' popularity-based listening histories and performances.
Particularly in exploration scenarios, lesser-known, unfamiliar songs might facilitate users' discovery of new content, improving their experience. To further investigate this phenomenon, we conducted a user study. Leveraging users' Spotify profiles, we generated personalized recommendations and applied mitigation algorithms. By means of questionnaires, we assess users' perceptions and satisfaction.
The results show no significant differences in satisfaction between the algorithms. While user-centred fairness goals seem not to influence the users' perception, reduced popularity achieved by an item-fairness-focused algorithm was perceived by the users, and the accompanying reduced Familiarity with the recommendations can enable Discovery, the feeling that the recommendations enrich the user's musical taste. Discovery is highly positively associated with satisfaction metrics and behavioural intentions. While no direct impact of the mitigation methods on behavioural intentions could be made, we show that high satisfaction predicts behavioural intentions.
This research advocates for mitigating biases in music recommender systems for the benefit of item providers (artists) and users, avoiding unfair treatment of distinct user groups. Moreover, we demonstrate the effectiveness of mitigation algorithms at an unprecedented scale and investigate real users' perceptions of the outcomes. Lastly, we explore users' behavioural motivations toward engaging with more equitable content. We posit that generating fairer recommendations can achieve a lasting influence on users' consumption behaviours, promoting an overall healthier music consumption pattern. | |
