Music recommenders show a strong gender bias
Music recommender algorithms such as those used on streaming platforms turn out to be strongly biased in gender. Christine Bauer, Assistant Professor in Information and Computing Sciences, found in a recent study that female artists count to no more than 25 per cent of all recommendations.
“Originally, my research colleagues and I were interested in what artists think about the way music recommenders work on streaming platforms,” says Christine Bauer, whose broader research topic is fairness of recommenders.
“Our impression was that the view of artists who are the core providers of content had not been researched yet. That is why we reached out to artists of different genres and with different popularity levels – some artists were just locally known, others were popular internationally - to see how they think of music streaming platforms; what they like and what they would like to have changed.”
The researchers chose for open questions, so as not to direct the artists’ answers to any particular direction. But what turned out, all nine artists we interviewed found gender as one of the main problems in music industry.
“Few female artists get attention, and this seems to be so over decades already. The artists see the recommender systems as a chance to move forward and to reach a better gender balance. I found it very interesting that the artists saw algorithms as a way to improve the world, whereas we are used to perceive them negatively.”
“So we looked at how music recommender algorithms perform with respect to gender, and we found that there are roughly speaking 25 percent female artists, and the same number goes also for the listener consumption: also there female artists make out just a quarter of the total. This imbalance is again reproduced by the algorithms. When recommendations are given, female artists count for 25 percent of the total. What is more, the female artists are lower down in the ranking.”
In their next step, the researchers ran a simulation that showed how the algorithms would behave over time. They saw that female artists remained at the lower end, and not only that: their percentage even decreased slightly.
“We then changed the algorithm a bit so as to see how it would behave in the simulation if users are listening to what is recommended. That is how it usually works: we consume what is recommended. Actually it is a cycle: what we consume is input for the recommendation and what the recommender recommends is what we consume. This helps understand how it comes that the percentage of female artists has remained constant for a long time.”
The vicious recommender-listening circle considered, it may not come as a surprise that Christine Bauer and her colleagues could see an immediate increase in the number of recommended female artists even at a slight change of the algorithm. The algorithm as it were learned to reproduce the new behaviour.
The platforms have a lot of possibilities to improve the gender balance in the algorithms.
However, she warns for drawing any definitive conclusions from the simulated situation, as it would need to be seen in practice whether the manipulated algorithm works that way. Although the researchers used a vast data set originating from LastFM, a platform that collects listening behaviour from various streaming platforms, the real test would be to have streaming platforms change their algorithms in reality.
“To summarize, we see that the platforms have a lot of possibilities to improve the gender balance in the algorithms, but also that they do have a lot of power.”
Music platforms and their recommender algorithms do not seem an obvious research choice for a computer scientist, but it does for Christine Bauer, who plays saxophone herself and joined a Utrecht-based big band recently.
“A lot comes together in this topic. For a long time already, I have done research on systems that adapt to persons and situations in different fields. Next to it, I am very much into music. I play and listen to music intensively and am befriended with many musicians. And before starting my academic career, I worked on copyright and licensing issues at a collecting society.”
“I find it especially important to reach out to the people who are affected by the way the algorithms work, the artists themselves. In the academia, we often make assumptions without finding out what the group studied really thinks.”
In a forthcoming article, she discusses another study on the use of voice in the voice assistant technology. And again, assumptions come into play.
“We wanted to find out which parameters make a voice sound more female or male. It is interesting that it is hardly possible to have a genderless voice, as people always perceive a gender in any voice they hear. I find this particularly interesting as some researchers have worked on creating a genderless voice. But another study shows clearly that there is no such thing: every single person in the study perceived a gender. This raises questions on how a voice is used in a voice assistant and what these voices implicate. My conclusion is that we should be careful in labelling the gender, as listeners differ in which gender they hear.”