How to break out of your Spotify feedback loop and find new music

Does the algorithm know you too well? Here’s how to shake up your recommendations for a more diverse listening experience
Spotify / WIRED

If you’re listening to music right now, chances are you didn’t choose what to put on yourself – you outsourced it to an algorithm. Such is the popularity of recommendation systems that we’ve come to rely on them to serve us what we want without us even having to ask, with music streaming services such as Spotify, Pandora and Deezer all using personalised systems to suggest playlists or tracks tailored to the user.

Generally, these systems are very good. The problem, for some, is that they’re perhaps really too good. They’ve figured out your taste, know exactly what you listen to, and recommend more of the same, until you’re stuck in an endless pit of ABBA recordings (just me?). But what if you want to break out of your usual routine and try something new? Can you train, or trick the algorithm into suggesting a more diverse range?

“That is tricky,” says Peter Knees, assistant professor at TU Wien. “Probably you have to steer it very directly into the direction that you already know you might be interested in.”

The problem only gets worse the more you rely on automated recommendations. “When you keep listening to the recommendations that are being made, you end up in that feedback loop, because you provide further evidence that this is the music you want to listen to, because you're listening to it,” Knees says. This provides positive reinforcement to the system, incentivising it to keep making more similar suggestions. To break out of that bubble, you’re going to need to quite explicitly listen to something different.

Companies such as Spotify are secretive about exactly how their recommendation systems work (and Spotify declined to comment on the specifics of its algorithm for this article), but Knees says we can assume most are heavily based on collaborative filtering, which makes predictions of what you might like based on what other people who have similar listening habits to you also like. You may think that your music taste is something very personal, but it’s likely not all that unique. A collaborative filtering system can build a picture of taste clusters – artists or tracks that appeal to the same group of people. Really, Knees says, this isn’t all that different to what we did before streaming services, when you might ask someone who liked some of the same bands as you for more recommendations. “This is just an algorithmically supported continuation of this idea,” he says.

The problem occurs when you want to get away from your usual genre, era or general taste and find something new. The system is not designed for this, so you’re going to have to put some effort in. “Frankly, the best solution would be to create a new account, and really train it on something very dissimilar,” says Markus Schedl, a professor at Johannes Kepler University Linz.

Failing that, you need to actively seek out something new. You could seek out a new genre, or use a tool outside of your main streaming service to find suggestions of artists or tracks and then search for them. Schedl suggests finding something you don’t listen to as much and starting a ‘Radio’ playlist – a feature in Spotify that creates a playlist based on a selected song (these may however also be influenced by your broader listening habits).

Knees suggests waiting for new releases or regularly listening to the most popular tracks. “There's a chance that the next thing that comes up is going to be your thing,” he says. But getting away from the mainstream is harder. You’ll find that even if you go actively searching for a new genre, you’ll likely be nudged towards more popular artists and tracks. This makes sense – if lots of people like something, it’s more likely you will too – but can make it hard to unearth hidden gems.

Knees advises therefore trying to actively dig into the “long tail” – the huge number of artists and tracks that have few listeners but might just be your niche. While you can manually trawl through obscure artists and back catalogues, however, your recommendations will likely still tend somewhat more towards the mainstream. “Even if you're in the long tail, it kind of pushes you back into the head, into the popular items, when making recommendations, because this is where the system is most stable,” he says.

As a general rule, if you want to diversify your listening, you’ll have to put more effort into music discovery rather than allowing the system to do it for you. Instead of just listening to personalised playlists, you could follow playlists curated by individuals, as well as making your own. “If you're relying on a platform to do the work for you, then you're basically in the radio mode, as people were before,” Knees says.

There is another way that music recommendation systems can work, which could help bust the feedback loop: content-based recommendations. In this approach, recommendations are based on sound rather than other people’s listening habits. The system could quantify aspects of music such as tempo and find similar tracks based on those acoustic qualities. Schedl suggests you could even put a numeric value on things like “danceability” or “instrumentalness”. In this case, you could even adjust the system for diversity, by tuning how similar recommended tracks should be.

How much this sort of content-based recommendation approach is used, however, is unknown, and it can be a very risky strategy in terms of user experience. Play too much of the same thing and a user might get bored; but play something too far out of their comfort zone and they might just leave.

“You have this trade off between sticking to really solid, no-risk recommendations by just doing what everybody does, and, on the other hand, letting the computer make a recommendation based on the sound properties alone without knowing anything about the cultural aspects of music, which might completely break that expectation,” Knees says. This could be good – it might find the perfect song just for you – or it could completely undermine a user’s trust in the recommendation system.

Meanwhile, if 2021 is the year you get back into music discovery, then you’ll have to take the initiative to explore outside of your filter bubble. It’s likely, in fact, that you listen to a greater range of music since using streaming platforms than you did before . Perhaps, muses Knees, it was the extra effort required to find an artist or track in the past that made it feel more precious. Put in the work, then, and it might just pay off.

Vicki Turk is WIRED's features editor. She tweets from @VickiTurk

This article was originally published by WIRED UK