The recommendation features of music-streaming platforms such as Pandora and Spotify have revolutionized how people find new songs and artists. But listeners with niche interests — particularly those who enjoy high-energy forms of rock and rap — are poorly served and "rarely receive relevant recommendations," a new study suggests.
The findings, published Monday in EPJ Data Science, could potentially improve music streaming algorithms for fans of more obscure music and give greater exposure to underground musicians, the researchers behind the study say.
"The idea is to have a little bit fairer music recommendation algorithms," said Elisabeth Lex, an associate professor at Graz University of Technology in Austria and a coauthor of the paper whose team studies music-streaming systems and user behavior. "I'm really passionate about not having just popular content being advertised. As a user, if you do not see that this kind of music even exists, you don't have a chance to explore."
As music listeners themselves, the researchers noticed a popularity bias in which the algorithms generating song and artist recommendations work well only for mainstream music.
"Many in our group like to listen to content that is not as popular," Lex told The Academic Times. "We all made this observation that we don't get good recommendations. It came out of personal curiosity — what is going on here?"
One possible explanation is that listeners of obscure subgenres don't create enough data to provide good recommendations. But in a study published last year, Lex's team disproved that notion, finding instead that those listeners have larger profile sizes and listen to a greater diversity of artists compared with users who engage with mainstream music. But algorithms still weren't targeting niche listeners accurately.
"Our hypothesis was that they have very curated taste," Lex said. "These are people who invest a lot of time in curating their music collections and do not profit as much from the social influence that comes with a recommendation system. Most work in a way that recommends music that similar users have listened to."
Using a publicly available dataset containing the listening histories of 4,148 users of the streaming platform Last.fm, the researchers identified 2,074 "beyond mainstream" music listeners.
Then they enriched the dataset with a clustering algorithm to sort nearly 3.5 million Spotify songs by their acoustic features. Songs were grouped according to characteristics such as energy, tempo, presence or lack of vocals, spoken-word elements and danceability.
Four large subgroups of non-mainstream music listeners emerged: acoustic folk; high-energy hard rock and hip-hop; ambient music with acoustic instruments and no vocals; and high energy music with no vocals, such as electronica.
All subgroups ranked low in terms of accuracy of recommendations, Lex said. The researchers found that ambient music lent itself more readily to the platforms' algorithms, while heavy rock and rap was the most difficult subgroup to classify.
They also examined the likelihood of listeners straying from their comfort zones.
"We found that people who belonged to the ambient music cluster also liked to listen to music from other classes," Lex said. "They're a little bit more open to music from other groups. Of course, that means it's easier to model their recommendations; they are easier to satisfy because they have a more open taste."
Listeners of aggressive, high-energy music were less open to other categories of music and "seemed a bit more individualistic than the others," Lex said. But they were more apt to engage with all different kinds of hard music, perhaps due to the existence of a multitude of extremely specific micro-genres of rap and rock.
"The genre taxonomy we used, it doesn't have just pop-rock," she said. "It has stuff like Viking death metal."
Moving forward, Lex and her colleagues intend to examine cultural and socioeconomic differences in how music is recommended on streaming platforms. They also plan on developing user models capable of accounting for the complexities of uncommon music tastes.
In addition to potentially benefiting consumers of niche music, the study's authors believe their work could help underground producers gain more exposure via streaming services. They also hope that streaming platforms will adjust their algorithms based on their research, Lex said. "From the platform's perspective, of course they want to have the best user experience for all their listeners."
The study, "Support the underground: characteristics of beyond-mainstream music listeners," published March 30 in EPJ Data Science, was authored by Dominik Kowald and Peter Muellner, Know-Center GmbH, Graz, Austria; Eva Zangerle, University of Innsbruck, Austria; Christine Bauer, Utrecht University, The Netherlands; Markus Schedl, Johannes Kepler University Linz, Linz Institute of Technology AI Lab, Austria; and Elisabeth Lex, Graz University of Technology, Austria.