Zoek in de site...

Music recommendations: And I still haven’t found what I’m looking for

Date of news: 4 March 2022

On music platforms, we have access to millions of songs and we are confronted with choice overload. What should we listen to next? Recommender systems —one of the most successful application areas of artificial intelligence — can help us navigate the music landscape: music recommenders suggest what we may enjoy listening to. Yet, if you like music beyond the mainstream, you might end up with poor recommendations. Also, artists are not represented equally well in the recommendation. Look at what your most preferred music platform recommends first. Chances are high that the first recommendation is a track by a male artist.

There is hardly anyone who does not listen to music. Have you ever managed to escape “Last Christmas” in the pre-Christmas season?! I bet you haven’t. Luckily, music platforms offer more than the tracks “Last Christmas” and “All I Want for Christmas Is You”. A lot more. On the major music streaming platforms, we can access more than 50 million tracks. So, what should we listen to next?

It seems like an easy question. But sometimes it is difficult to answer. What exactly should come next? Sometimes, we want to actively explore the deep blue sea of tracks. What’s hot at the moment? Wow, this “new” artist has released albums in the 1990s already?! How come that we have not come across these before?

Sometimes, we just want to lean back and listen. Or work and listen. Or do housekeeping and listen. In such situations, we focus on something else and don’t want to actively decide on each and every track to be played.

It is a relief that we have so-called recommender systems that can help us navigate the music landscape—getting us what we like, while not repeating the same tracks over and over again.

How do music recommender systems work?

One of the major approaches used in recommender systems, content-based filtering, considers the content (e.g. timbre, tempo, key, danceability, melody) of what you listened to and identifies similar tracks or artists. You get more of the same.

Another group of approaches ignores the content, but rather considers what other users like. For instance, a simple approach considers what is currently popular on a platform (i.e. what most users like) and recommends those items to you, which is a popularity-based approach. Why does it work? Well, there is a higher chance that a random person likes a song that is generally popular than that this very person likes a random unpopular song. A more promising approach is to watch out for “similar” users in terms of what they listened to and to recommend to you what those similar users liked (collaborative filtering).

For all, but not for everyone

In the ideal case, you get what you want. That’s great. In reality it does not work out equally well for all sorts of users. For instance, music listeners interested in mainstream music (thus, music that is widely popular) are traditionally served well by music recommender systems, whereas users interested in music beyond the mainstream rarely receive relevant recommendations. However, listeners in some countries follow the global mainstream, while other countries (also) have their country-specific mainstream. Thus, recommenders do not work equally well across countries. Also, sub-groups of beyond-mainstream listeners experience different outcomes. Recommendation algorithms seem to be worse at picking tracks for fans of hard music, who are not much interested in music outside of their niche, than for ambient music lovers, who seem happy to listen to artists outside of their niche.

What about the artists?

Now let’s imagine being an artist. You want to conquer the world. What do you need to do? Well, besides tuning your guitar and practice, there are also other factors involved. Recommenders influence who gets exposure — who gets attention and whose songs get recommended even more. And that might affect popularity — making a living, making a career.

We reached out to music artists to understand how they feel affected by current streaming platforms—such as Spotify, Pandora, or Tidal—and what should be improved. It did not come as a surprise that they saw some problems. For instance, they don’t feel in control of how their work is presented. One artist states that some platforms prominently present his early-career work: “[…] it is something that I have done 10 years ago. [The platform] puts the most listened tracks [at the top]. […] and you have to scroll down to reach the latest album”.

The same old story

Artists also see room for improvement for recommender systems. From the interviews, we understand that one of the major problems that the artists see in the music business is the gender imbalance, which has persisted for decades.

Artists see recommenders as a way to reach gender balance. One artist states: “I think there should be actions to correct some biases.”

It did not come as a surprise that the analysis of users’ listening behavior over several years (for datasets, see) showed that only 25% of the artists represented were female.

From this imbalanced starting point, we analyzed the performance of a widely-used collaborative filtering technique. For each user in our dataset, we computed personalized music recommendations in the form of a ranked item list (as it is typically done). The proportion of female and male artists is similar to the general listening behavior (~25%). Yet, on average, the first track by a female artist ranks on the 7th or 8th position; the first track by a male artist, on average, comes on position one.

As users listen to what they are recommended, and in turn, the recommender learns what the users listen to, this creates a feedback loop which can amplify the gender imbalance. As a result, the exposure of female artists remains low.

What if we tweak the algorithm a bit? Can we break it?—Yes, we can!

We propose a simple approach to gradually give more exposure to female artists: we take the recommendations computed by the basic recommendation approach, and then we re-rank the output by moving male artists a specified number of positions downwards. By simulating the feedback loop, we show that gender can be better balanced in the longer term by gradually increasing the exposure of female artists in the recommendations. Most notably, this balance is achieved without severely affecting the recommender’s performance—thus, the system still recommends tracks that the users like.

Our analysis shows that once the users start changing their behavior, the recommender starts recommending more females over time. In other words, we break the feedback loop.

We’ve got work to do

Our computational simulation is promising. Now, can we put this into practice? I am convinced we can. But, well, I am not planning a university spinoff with the next music streaming service. So, oh you streaming platforms in the world, please pick up this approach. You can help change the world.

While we have one potential solution in place, we are not done. What about non-binary people? We do not even have data—yet! What about imbalances between countries? Regions? Across genres? Combinations of those? Surely, we’ve got work to do.

Turning Magazine is looking for a new board!

Turning is a student-run platform where professionals, students, and companies can write and read about recent developments in AI. Are you a student, or a group of students, with fresh ideas and passionate about bridging the gap between AI research and society? Join Turning's team and become a board member: chair, secretary, editor in chief, designer, social media manager or external affairs officer.

What Turning offers you:

  • weekly fun meetings
  • you help create a biannual magazine that is read internationally
  • expand your network
  • increase your teamwork skills and make new friends
  • spend approximately 1 to 3 hours a week on Turning

Interested or want to know more? Find us on Instagram or LinkedIn and send an email to info@turningmagazine.com.