Nowadays more and more people are listening to music on streaming apps – by early 2020, 400 million people were subscribed to one. These platforms use algorithms to recommend music based on listening habits. Recommended songs can appear in new playlists or they can start playing automatically when another playlist is finished.
But what algorithms recommend isn’t always right. In a new study, we have shown that a widely used recommendation algorithm is more likely to choose music from male artists than from female artists. In response, we found a simple way to give more visibility to female artists.
The representation of women and gender minorities in the music industry is extremely low. About 23% of the 2019 Billboard 100 artists were women or gender minorities. Women make up 20% or less of recorded composers and songwriters, while 98% of works performed by large orchestras are by male composers.
This bias is also present in streaming services. A few female “superstars” dominate among the most popular artists, but most female and mixed artists are in the lower popularity levels. While the problem extends beyond the music industry, online music platforms and their algorithms that recommend music – called recommenders – play an important role.
Read more: Music streaming: Listening to playlists lowers the income of small artists
While previous studies have repeatedly asked consumers for their opinion, musical artists, those who provide the content, are seldom aware.
We wanted to highlight the artists. We asked musicians for their thoughts on what would make online music platforms more fair. When they said that the gender imbalance was a major problem, we decided to study it in more detail.
Our analysis of the listening behavior of around 330,000 users over nine years showed a clear picture – only 25% of artists ever listened to were female. When we tested the algorithm, we found that on average the first recommended lead was male, with the next six. Users had to wait until the seventh or eighth song to hear one from a female.
Breaking the loop
As users listen to the recommended songs, the algorithm learns. This creates a feedback loop.
To break this feedback loop, we have proposed a simple approach to gradually give more visibility to female artists. We took the recommendations calculated by the base algorithm and reclassified them – moving male performers down a specified number of positions.
In a simulation, we investigated how our reclassified recommendations might affect users’ listening behavior in the long run. With the help of our reclassified algorithm, users would start to change their behavior. They listened to more female artists than before.
Eventually, the recommender began to learn from this change in behavior. He started placing women higher on the recommended list, even before our reclassification. In other words, we have broken the feedback loop.
It shows how easy it can be. Our simple method can help resolve algorithm biases that play an important role in how many people discover new music and artists. Next, we hope to study how real consumers perceive the changes introduced by the reclassification strategy and how it impacts their long-term listening behavior.
Another crucial step would be to collect and use data on the large scale of gender identities. We are aware that this binary gender classification does not reflect the multitude of gender identities. The unavailability of data beyond the gender binary is a major obstacle, both for research and for taking action and making progress at the societal level.
So far, our simulation could demonstrate the advantages of a simple reclassification approach. But of course, the responsibility does not lie solely with the platform providers. Initiatives such as Keychange and Women in Music strive to represent underrepresented people in the music industry. The rest of us must follow.
While music festivals are criticized for the lack of women in their rosters, any step towards a more balanced representation of more women of all genders is a step in the right direction.
This article by Christine Bauer, Assistant Professor of Human-Centered Informatics, University of Utrecht and Andrés Ferraro, PhD Student, Information and Communication Technologies, Universitat Pompeu Fabra, is republished from The Conversation under a Creative Commons license . Read the original article.
Published March 31, 2021 – 13:00 UTC