Ir al contenido

Documat


Resumen de Music recommender systems: taking into account the artists’ perspective

Andres Ferraro Paolino

  • Music streaming platforms nowadays play an important role in music consumption and have a big influence on the musical taste of the listeners. Machine learning-based recommender systems are a fundamental part of such streaming platforms defining what music people listen to and when. As for many applications of machine learning, there is an increasing debate in academia, industry and governments about the effects that recommender systems have in society and the ethical implications of such systems.

    Bias in music recommender systems towards more popular items has been studied extensively in the past. This bias affects both artists and listeners since it reduces the possibility of a large catalog's portion of getting any exposure. Recently, in the recommender systems community, it was raised the importance of considering the multiple stakeholders of a system when generating the recommendations. However, most of the research in the music domain has taken into account the users' perspective only. This thesis goes beyond the problem of popularity bias, it tries to uncover other dimensions in which the music recommender systems can affect the artists and propose alternatives to mitigate such problems.

    The contributions of this thesis are (i) identification of multiple aspects in which the current platforms and their recommender systems affect the music artists and concrete ways in which they could be more beneficial in the future, (ii) analysis of the algorithmic effect regarding gender imbalance in the recommendations and mitigation of such problem based on the output of artists' interview, (iii) analysis of the longitudinal effect of multiple state-of-the-art algorithms for session-based recommendations in users behavior negatively affecting the artists, (iv) publication of the first large-scale open dataset that contains audio and playlist information, (iv) novel contrastive learning approach proposed to combine multiple modalities (audio, genre and playlist information) beneficial for multiple tasks such as music recommendation, genre classification and automatic-tagging.

    It is necessary to improve recommender systems through multidisciplinary research. Contributions like the ones presented in this thesis allow us to move a step forward in that direction, making streaming platforms more beneficial for both the artists and users.


Fundación Dialnet

Mi Documat