Algorithm-Generated Music Recommendations: Low Accuracy for Fans of Beyond-Mainstream Music

A new study published in EPJ Data Science shows that fans of beyond-mainstream music, such as hard rock and ambient, may receive less accurate recommendations than fans of mainstream music, such as pop. A team of researchers at Know-Center Graz, Graz University of Technology, University of Innsbruck, University of Utrecht, and Johannes Kepler University Linz analysed the characteristics of beyond-mainstream music listeners to better understand the needs of this underserved user group.

Popularity Bias in Music Recommendations

As more and more music is available via music streaming services, music recommendation systems have become essential for helping users to search, sort and filter extensive music collections.

Beyond-mainstream listeners receive worse music recommendations than mainstream listeners

However, it is a widely-known problem that recommender systems are prone to popularity bias, which leads to long-tail items (i.e., items with few user interactions) having little chance of being recommended.

We validated this effect by measuring the accuracy of algorithm-generated music recommendations for listeners of mainstream and beyond-mainstream music.

We used a dataset containing the listening histories of 4,148 users (2,074 users in each group) of the music streaming platform Last.fm who listened mostly to beyond-mainstream music (BeyMS) or mostly to mainstream music (MS).

Figure 1 shows that beyond-mainstream listeners receive worse music recommendations than mainstream listeners.

Figure 1. Recommendation accuracy measured by the mean absolute error (lower is better) of a non-negative matrix factorization-based approach (NMF) and a neighborhood-based approach (UserKNN) for mainstream and beyond-mainstream users
© The Authors

Subgroups of Beyond-Mainstream Music Listeners

We applied the unsupervised clustering algorithm HDBSCAN* to identify subgroups within the beyond-mainstream music listeners.

We identified four subgroups, which we labeled according to the types of music they most frequently listened to: (i) users of music genres with only acoustic instruments such as folk (U_folk), (ii) users of high-energy music such as hard rock or hip-hop (U_hard), (iii) users of music with acoustic instruments and (nearly) no vocals such as ambient (U_ambi), and (iv) users of high-energy music with (nearly) no vocals such as electronica (U_elec).

Impact on Music Recommendations

the willingness of users to listen to music outside their own music preferences has a positive effect on the quality of music recommendations

By comparing each subgroup’s listening histories, we identified users who were most likely to listen to music outside their preferred genres.

Those who mainly listened to acoustic music with (nearly) no vocals such as ambient (U_ambi) were found to be most likely to listen to music preferred by the other subgroups as well.

Those who mainly listened to high-energy music such as hard rock or hip-hop (U_hard) were least likely to listen to music preferred by the other subgroups.

In figure 2 below we see that U_ambi receives better recommendations than U_hard, which means that the willingness of users to listen to music outside their own main music preferences has a positive effect on the quality of music recommendations.

Figure 2. Comparison of the mean absolute error scores reached by non-negative matrix factorization for the four beyond-mainstream subgroups with the ones reached for the whole beyond-mainstream group (BeyMS) and mainstream group (MS)
© The Authors

Towards Fair Music Recommendations

We believe that our findings provide valuable insights for developing improved user models and recommendation approaches to better serve beyond-mainstream music listeners.

However, we also believe that a lot of research is still needed to provide fair music recommendation models that are generalizable and avoid the unfair treatment of any user group.

We hope that our data set (https://doi.org/10.5281/zenodo.3784764) and source code (https://github.com/pmuellner/supporttheunderground) provided with the article are of use to the scientific community for future analyses.

View the latest posts on the On Physical Sciences homepage

Comments