BIASES IN MUSICAL RECOMMENDER SYSTEMS: SURVEY, FORMALIZATION, CLASSIFICATION AND DEBIASING METHODS

Seniatkin Nikita Stanislavovich

The paper provides a structured systematization of biases inherent to recommender systems with aim to musical domain. The survey introduces an integrated taxonomy that captures both general algorithmic distortions and domain-specific effects characteristic of the music industry. The study formulates the core objective of consolidating fragmented knowledge about data-driven, algorithmic, and socio-contextual biases into a unified analytical framework. It identifies key distortions observed in practice, including popularity, exposure, and position biases, as well as genre, cultural, gender, and label-related inequalities amplified by feedback loops and repetitive content consumption behavior. The work formalizes these phenomena using quantitative measures such as popularity and diversity indicators, exposure and fairness oriented metrics, and bias amplification coefficients. It further classifies mitigation strategies applied at the stages of data preprocessing, model training, and post-ranking adjustment, emphasizing the need for domain adaptation due to the strong impact of playlists, seasonal trends, and structural asymmetries between artists. The results establish a methodological foundation for evaluating fairness in music recommendation and outline research prospects, including the development of domain-specific metrics, datasets enrichment, and frameworks that find fairness balance between users and artists interests.

Keywords:recommender systems; music streaming; music recommender systems; machine learning; data bias; bias evaluation; genre bias; specific bias; debiasing