(Credit: cottonbro from Pexels)
BARCELONA, Spain — Critics often accuse the music industry of placing men on a pedestal, while failing to give the same respect to female artists. Now, a new study reports even song-suggesting computer algorithms are far more likely to recommend hits from male performers.
Initially, researchers from Universitat Pompeu Fabra in Spain just wanted to investigate the fairness of various online music platforms from the actual artist's perspective. However, they soon discovered gender equality is still a big concern among most musicians, male and female alike.
Study authors tested a popular music recommendation algorithm across two song datasets. That process revealed that in both instances the algorithm ended up “reproducing” existing biases in the datasets. Results reveal both datasets displayed biases against women, with female artists only accounting for 25 percent of the songs.
Algorithms keeping female singers from topping the charts?
Moreover, the algorithm always constructs a ranked list of songs for a user to sample. Across the board, the highest a female artist ever ranked on such lists was number six or seven.
“The bias in exposure comes from the way recommendations are generated,” says study author Andrés Ferraro in a university release.
In other words, female artists aren't gaining nearly as much exposure as their male counterparts.
Adding it this issue, researchers find most people usually end up listening to songs their playlists suggest to them. So, as an individual listens to more suggested songs, that reinforces the idea that the algorithm is doing something right, creating a “feedback loop” of gender bias.
To fix all this, researchers suggest manually reordering song recommendations to ensure more female songs are suggested earlier. A trial using this adjusted algorithm even appears to correct the issue. Over time, users using the new algorithm started listening to more female singers.
The study appears in Proceedings of the 2021 Conference on Human Information Interaction and Retrieval.
