The most used applications for listening to music, such as Spotify, Last.fm or own Youtube, have algorithms capable of predicting and showing you new music that you might like. In a simple way, it is a recommendation system through collaborative filtering: the apps register the artists and genres that a user listens to and compare these results with like-minded listeners to know what others like.
Thus, Lil Nas X lovers are recommended, quite rightly, to listen to Post Malone or, if you liked Soleá Morente, they will have recommended you Rigoberta Bandini.
A computer model based on music history predicted whether users would like a recommended song using four different algorithms
But these algorithms are not perfect with something as subjective and human as artistic creation and musical tastes. For this reason, a team of researchers from the Graz University of Technology, the Know-Center GmbH research center, the Johannes Kepler University of Linz, the University of Innsbruck (all from Austria) and the University of Utrecht (Netherlands) have wanted test how accurate are the recommendations generated by these algorithms, especially for listeners of music that is not very popular or not so well known to the general public. The main result, published in the latest issue of the journal EPJ Data Science, is that these algorithms they fail a lot more in listeners of hard rock and hip-hop, than with other musical genres.
To verify this, the team took the history of songs listened to by 4,148 users of the Last.fm platform, both from listeners who usually listen to more popular commercial music, and those who prefer somewhat less well-known artists (2,074 users in each group). Based on the artists most listened to by each user, the research used a computational model to predict if they would like a new song or artist using four different recommendation algorithms. In this way, they confirmed that popular music listeners tend to receive more accurate and precise recommendations than the group of less commercial listeners.
Acoustics listeners, within the unpopular music group, received the best recommendations. On the other hand, those of energetic, like hip-hop and hardcore, received the worst
Following this, the authors categorized non-commercial music listeners into four groups, according to the characteristics of the music they most often listen to. These groups were: listeners of musical genres containing only instruments acoustic, like folk or singer-songwriters; music very energetic like punk or hip-hop; very music acoustic but no voice like ambient music; and music very energetic but no voice like electronics. The research was thus able to compare the histories of each group and identify, with the computational model, which users were more likely to listen to music outside their preferences and the diversity of musical genres within each group.
Thanks to this categorization, the study found that listeners of acoustic music without voice also tended to prefer songs from the other three groups (energetic, energetic without voice, and acoustic) and received more accurate recommendations from the computational model. At the same time, the group of listeners of energetic music were the ones who received the worst recommendations algorithms, despite the fact that his group featured the widest variety of music genres – hard rock, punk, hardcore, hip-hop, and pop rock.
Popularity bias in algorithms
Elisabeth lex, co-author of the work and associate professor of applied informatics at the Graz University of Technology, highlights that music recommendation algorithms are already “essential” for users who want to search, select and filter the collections of music applications.
The analysis is based on a sample of Last.fm users, which could be unrepresentative for this or other music platforms
Despite this, it indicates that algorithms may fail to make recommendations for non-commercial music listeners. “This may be because these systems are biased towards popular music, getting artists outside the mainstream to be less listened to ”, he points out.
Finally, the authors suggest that their findings could serve as the basis for creating music recommendation systems that provide more accurate recommendations. However, they warn that their analysis is based on a sample of Last.fm users, which could be unrepresentative for this or other music platforms.
Kowald et al. “Support the underground: characteristics of beyond-mainstream music listeners”. EPJ Data Science (2021). DOI: 10.1140 / epjds / s13688-021-00268-9
Rights: Creative Commons.