Take a fresh look at your lifestyle.

Music recommendation algorithms reveal gender bias

A study by the Music Technology Group (MTG) in Barcelona has shown that a commonly used recommendation algorithm is more likely to choose music from male performers at the expense of female performers.

While the problem of gender discrimination is already thriving in the music industry, the study by researchers from Pompeu Fabra University (UPF), Barcelona and Utrecht University (UU), Netherlands, found that music recommendation algorithms widen the gender gap.

Andrés Ferraro and Xavier Serra, the MTG researchers at UPF, together with Christine Bauer from UU, recently published a paper on gender balance in music recommendation systems in which they wonder how the system should work to prevent gender bias.

Initially, the work of Ferraro, Serra and Bauer was meant to understand the fairness of music platforms available online from the point of view of the artists. In interviews with music artists, they found that gender justice was one of their main concerns.

The team tested a widely used music recommendation algorithm based on collaborative filtering and analyzed the results of two data sets.

In both cases, they found that the algorithm reproduces the existing bias in the dataset, in which only 25 percent of the performers are female.

In addition, the algorithm generates a ranking of artists to recommend to the user. Here, the researchers found that, on average, a female performer’s first recommendation comes in the 6th or 7th position, while that of a male performer is in the first position.

“The bias in exposure comes from the way it generated recommendations,” Ferraro explained, meaning that women are less exposed based on the system’s recommendations.

The researchers said the situation worsens when they consider that the algorithm learns when users listen to recommended songs.This, in turn, creates a feedback loop.

But with the help of the reordered algorithm, users change their behavior so that they listen to more female performers.

The researchers have proposed an alternative approach that would allow greater exposure of female performers and would consist of rearranging the recommendation, which would move a certain number of positions down to resolve the existing gender bias.

In a simulation, the team examined how classified recommendations would affect user behavior in the long term. The results showed that using the reclassified algorithm, users would change their behavior and thus listen to more female performers than with other music recommendation algorithms, and the new algorithm, based on machine learning, would merge this behavioral change.