Humans + Machines: A Look Behind the Playlists Powered by Spotify’s Algotorial Technology
TL;DR Since 2017, Spotify has been working to create a better listening experience for our users by creating algorithmically personalized playlists powered by the expertise of our curators. The outcome of these efforts has resulted in the technology we call “Algotorial.”
The best of both worlds: Editorial and Algorithms
Spotify has a library of playlists for almost every occasion, be it different moods, activities, genres, eras, new music, top hits, emerging artists, cultural moments, or regional listening trends.
Some of these playlists, such as RapCaviar, created and maintained by our Editorial team, are distinctive based on genres or localities. One editor might concentrate on Hip-Hop while another editor focuses on the music being produced and listened to in Brazil. This specialization allows them to have a deeper understanding of the music that artists have created and the ways users listen to them. This is particularly powerful as it allows editors to be aware of small trends or cultural events that drive the ever-changing way users are consuming music.
While the RapCaviar playlist is owned by our Editorial team, others like Discover Weekly, Daily Mix, and Your Time Capsule are powered by our Personalization Algorithms. These algorithms take a look at the audio attributes of music and can find similarities across tracks to identify what songs are often listened to together. So by combining a user’s listening history with the relationships across tracks, we can create a unique list of songs for each one of our users, based on what we think they will like.
Creating a playlist with Algotorial technology
The process of creating Personalized Editorial Playlists starts with the editors. Our editors begin by envisioning a specific user need — let’s take an activity like a road trip, for example. Once the user’s need is defined, the editor creates a content hypothesis — in our road trip example, the editor decides the content hypothesis could be “familiar songs you know all the words to, and would sing along to”.
Knowing what songs might be particularly “singable” is difficult to describe in algorithms. It might be the song that was on repeat last summer, or a particularly catchy refrain that is repeated over and over. Maybe it appeared in a TV show or movie recently to remind you of your teenage years. It’s hard to describe why, but you know it when you hear it. This is where human intuition comes in.
The editor collects tracks that could be appropriate for the playlist and adds them to what we call a “pool”. This large pool combines their musical and cultural expertise, as well as advanced filtering for searches they’ve conducted to help narrow down the vast universe of possible songs to the most relevant ones. They also have access to metrics to see what tracks have been performing well in the playlist and which have not.
Because the editors are choosing potential candidates instead of the exact order of the final playlist, they can expand the pools to a wider range of tastes and not just the most obvious and popular tracks. They don’t need to aim for a balance that will make everyone happy. They can pick songs that will appeal to a wide range of listeners.
After the pool is created, the algorithms take over, picking out the appropriate tracks and placing them in order for each given user. This is particularly helpful for playlists on broad topics. In our road trip example, the playlist might have a mix of Pop, Indie, Rock, and Hip-Hop in the pool. With personalization, this one playlist can suit a wide variety of listeners while trusting that all of the candidates are still “singable”.
Finally, the editor works with the Spotify design team to brand the playlist’s title, description, and imagery. They can further personalize by having the imagery optionally chosen based on the listener’s tastes. For example, a ’60s Rock playlist might have one image with a British Invasion artist and another with a Surf Rock artist. We will display the image of the artist for which the user has the most affinity.
The end result of these combined efforts? A new Algotorial-powered, personalized editorial playlist designed for road trips: Songs to Sing in the Car.
Adapting to our listeners
In broad strokes, we use various machine learning techniques to analyze a user’s listening history to better predict which songs they will want to listen to. We then take those preferences and apply an order to the tracks in a way that flows together, creating an enjoyable listening session.
As listeners engage with the playlist, their actions such as listening, skipping, or saving to their library help train our recommendation engine about how best to use the tracks in our music library. Additionally, those signals influence our representation of the listener’s taste profile to improve the recommendations they receive in the future. We are simultaneously learning ways to improve our recommendations for all users as well as for the individual listener.
Learning from experts
Personalization is at the heart of what we do. When we ask our listeners what they like most about Spotify, more than 81% cite our personalization. Personalized Editorial Playlists has been extremely successful in taking the expertise of our editorial team and scaling it, so that every listener on Spotify can have a personalized experience. Through this collaboration between humans and machines, we are constantly learning and improving what goes into a great listening experience. This allows our passion for music to be shared with millions of users every day.
Looking for a playlist for a specific occasion? Check out some of our favorite Algotorial-powered playlists, including the aforementioned ones, on the Spotify app:
- Locked In
- lofi beats
- I Love my ’90s Hip-Hop
- 90s Country
- Mood Booster
- Songs to Sing in the Shower
- Songs to Sing in the Car
- Beast Mode
- Chillin’ on a Dirt Road
- my life is a movie
- Classic Road Trip Songs
Tags: engineering leadership