Assessing the impact of music recommendation diversity on listeners
- Autores: Lorenzo Porcaro
- Directores de la Tesis: Emilia Gómez Gutiérrez (dir. tes.), Carlos Castillo Ocaranza (codir. tes.)
- Lectura: En la Universitat Pompeu Fabra ( España ) en 2022
- Idioma: español
- Tribunal Calificador de la Tesis: Christine Bauer (presid.), Perfecto Herrera Boyer (secret.), Mounia Lalmas Roelleke (voc.)
- Programa de doctorado: Programa de Doctorado en Tecnologías de la Información y las Comunicaciones por la Universidad Pompeu Fabra
- Materias:
- Enlaces
- Tesis en acceso abierto en: TDX
- Resumen
- español
Esta tesis se centra en evaluar el impacto que la diversidad en las recomendaciones musicales puede tener en los oyentes. En el ámbito musical, la diversidad es uno de los valores que los sistemas de recomendación deben preservar, ya que el patrimonio musical mundial es una mezcla de varios lenguajes artísticos y paisajes sonoros, y las diferencias están en el centro de esos procesos de fusión. Sin embargo, una corriente de estudios críticos ha sacado a la luz varios problemas debidos al uso de los sistemas de recomendación, en la raíz de los fenómenos como el empeoramiento del sesgo por popularidad, la discriminación hacia grupos históricamente infrarrepresentados en la industria musical o el refuerzo de hábitos de escucha homogéneos. Explorando la medición, la percepción y, finalmente, el impacto de la diversidad, discutimos cómo favoreciendo la exposición a música diversa, las recomendaciones algorítmicas pueden ayudar a las personas a entender su Yo musical observando las "otras" culturas con las que interactúan.
- català
Aquesta tesi se centra a avaluar l’impacte que la diversitat en les recomanacions musicals pot tenir en els oients. En l’àmbit musical, la diversitat és un dels valors que els sistemes de recomanació han de preservar, ja que el patrimoni musical mundial és una barreja de diversos llenguatges artístics i paisatges sonors, i les diferències són al centre d'aquests processos de fusió. Tot i això, un corrent d'estudis crítics ha revelat diversos problemes deguts a l'ús dels sistemes de recomanació, a l’arrel dels fenòmens com l’empitjorament del biaix per popularitat, la discriminació cap a grups històricament infrarepresentats a la indústria musical o el reforç d’hàbits d’escolta homogenis. Explorant la mesura, la percepció i, finalment, l'impacte de la diversitat, discutim com afavorint l'exposició a música diversa, les recomanacions algorítmiques poden ajudar les persones a entendre el seu Jo musical observant les “altres” cultures amb què interactuen.
- English
The act of listening to music in today’s digital spaces is highly reliant on streaming services. In the last two decades, these platforms have radically changed how we search, interact, and ultimately enjoy music. Among the technologies behind the great success of such a paradigm shift in listening practices, Music Recommender Systems (Music RS) are one of the most influential tools, guiding users in finding the music that best fits their demands. Hidden behind streaming services’ home pages or explicitly presented in the form of “Made for you”-like playlists, algorithmic recommendations help people in overcoming the overwhelming flow of information contained in contemporary platforms, acting as a filter of the enormous catalogue of music accessible online.
Apart from several benefits, a strain of critical studies has brought to light several issues due to the ubiquitous use of recommender systems. Phenomena such as the exacerbation of the popularity bias (e.g. long-tail problem), discrimination towards historically underrepresented groups in the music industry (e.g. gender bias), or the reinforcement of homogenous listening habits and behaviours at the expense of the possibility to explore diverse content (e.g. filter bubbles), made evident the disparate impact that Music RS are having on listeners and artists.
This thesis focuses on assessing the impact that music recommendation diversity may have on listeners. In the music domain, diversity is one of the values that recommender systems should preserve, because the world music heritage is a mixture of different artistic languages and sonic landscapes, and differences are at the heart of such processes of melting. By exploring the measurement, perception and finally the impact of diversity, we discuss how favouring the exposure to diverse content, algorithmic recommendations may help people in understanding their musical Self by observing the "other" cultures with which they interact.
This work is divided into four main sections. In Section 2, we conceptualise diversity by reviewing the existing literature in 1) Information Technology, and in particular Information Retrieval and Recommender Systems literature; 2) Human-Computer Interaction, with a focus on Social Computing; 3) Social science, specifically Media Studies and Music Sociology. The interdisciplinary scope of this section is dictated by the multifaceted nature of the concept of diversity, a scope which is maintained through all the work at hand.
Afterwards, we introduce in Section 3 two indexes to measure the diversity of a music list. The first index is computed from the tracks’ popularity, the second one is based on the semantic information contained in user-generated tags. The advantages and disadvantages of such system-centric and algorithmic-driven metrics are discussed, in particular evidencing how the lack of feedback from listeners, the end-users of Music RS, strongly limits the validity of those measurements.
We explore in-depth the listeners’ perceptions of diversity in Section 4, wherein we report the results of a user study in which we ask people to indicate which among two lists, composed by tracks, artists and tracks and artist together, is the most diverse according to their opinion and beliefs. We then compare participants’ responses with the output of a computational model designed to identify the most diverse lists, according to tracks’ audio features and artists’ characteristics. By means of agreement analysis, we find that computational models are generally aligned with participants' choices when most of them agree that one list is more diverse than the other. In addition, we observe how differences in domain knowledge, familiarity, and demographics influence the level of agreement among listeners, and between listeners and computational diversity metrics. We complement the results of the user study by interviewing several participants about their experiences while participating in the study, and more in general about the role that music recommendation diversity has in their daily life.
Lastly, to assess the impact of music recommendation diversity we design a three months longitudinal study wherein we expose listeners to music recommendations with different levels of diversity, selecting tracks and artists from a musical genre unfamiliar and unknown to them. The familiarity inclusion criterion has been chosen based on our previous user study, wherein we show how participants who do not have a deep knowledge of a music genre, contrary to experts, tend to agree more among themselves, having a vision of diversity less influenced by previous experiences, prejudices or preferences. Our main hypothesis is that participants receiving recommendations with a higher degree of diversity will be more likely to listen and to appreciate unfamiliar music, in comparison to those who are exposed to a music unknown and less diverse. The results of this study are presented in Section 5.
- español
