How does Spotify recommend music to its users?
In 2015, Spotify launched its very first “Discover Weekly” playlists to its users and it became an instant hit. Within 6 months of launch, this playlist was streamed over 1.7 billion times, according to the company.
Nowadays, we can see a few more playlists or collections uniquely curated for each user, such as “Trending genres for you”, “Uniquely yours” (collection), “Based on your recent listening”, “More Like [Artist Name]”, and “More of what you like”. Its recommendation system covers not only music but podcasts as well, such as a curated list of “Shows you might like”. At the time of writing, my Spotify homepage has 21 “component positions” vertically, and 8 of them directly indicate music recommendations, such as “More Like [Artist Name]”. And for the rest, nearly all 13 of them have embedded some sort of recommending mechanism into the collections in their positions with the exception of “The state of music today”, which I assume is intended to introduce users to new genres and music, with which the users might discover new music tastes either similar to or unlike what they were used to before.
But how does Spotify recommend music (50 millions+ tracks, 1 million+ podcasts) to users? Why is the recommendation system so important to Spotify? How do companies like Spotify figure out if its recommendation system is working well or not? What are the key metrics they look at when evaluating recommendation performance?
How does Spotify recommend music to users?
The “Discovery Weekly” algorithms look at two fundamental pieces of information. First, it looks at what are the songs that you have listened to and liked or added to your playlists; Second, it looks at the songs in other people’s playlists. They leverage these two basic pieces of information and decide what songs to recommend to a specific user. For example, if someone has 5 songs in one of their playlists and you have 4 of them in your library, then chances are you are also going to like the fifth song that’s in their playlist. And that song will therefore appear in your “Discovery Weekly”.
This technique is called “Collaborative Filtering”, which is widely applicable to other services as well. For example, Amazon uses this technique to recommend items by looking at your purchase history and that of others. Netflix uses this to recommend shows and movies based on your watch history and other people’s watch history. Other examples include Youtube’s video recommendations, Facebook’s friend suggestions, Uber Eats’ restaurant suggestions, etc. And the algorithm works even better when there is a large user base as collaborative filtering gets more useful when there are more users. However, it is worth noting that with more users and data, it will require more computing power and storage, which can be a big scalability challenge as their user base grows.
One interesting assumption embedded in this technique is that people’s playlists are all connected in some way. A visual representation could be a huge network of millions of dots with each dot representing a playlist. Some playlists connect to a lot of other playlists and some only connect to 20 other playlists or less. But no matter how eccentric and unique a musician is, Spotify can still find ways to connect them to the right audience that likes their style of music.
To really resonate with users like a good friend, Spotify uses a second method called “taste profile” to recommend songs to its users. They profile the users’ music tastes, based on their listening behavior and habits, by generating and grouping clusters of artists, genres, and micro-genres — not just as broad as “pop” or “jazz” but as fine-grained as “synthpop”, “southern souls”, and “New Americana”.
Strategic Analysis
1. Spotify’s mission
Unlock the potential of human creativity – by giving a million creative artists the opportunity to live off their art and billions of fans the opportunity to enjoy and be inspired by it.
2. Spotify’s market share
According to Statista, a global business data platform, Spotify has about 30% of the market share worldwide in the music streaming industry, the most out of all music providers.
3. Spotify’s competitors
Competitors include Apple, Amazon, Google (YouTube), Pandora, etc.
4. Spotify’s users
At the time of writing, around 30% of listeners are Millennials, with 25% being Gen Z. Interestingly, 19% of users are above the age of 55.
As of March 2020, Spotify offers 4 billion playlists for users to listen to in 79 countries and has 286 million monthly active users (MAU), including 113 million Premium subscribers.
As of March 2020, Spotify has more than 50 million tracks, 1 million podcasts, and 4 billion playlists.
A large user base and a large content library are crucial for Spotify to carve out a niche and differentiate itself from its competitors.
5. Spotify’s financials
- Revenue 2019: ~$6.76 billion with Premium subscribers accounted for about 90% of its revenue.
- Business model: 30-day free trial, Ad-based free tier, premium plans (regular, family, student); with some plans, it is bundled up with other companies/services such as Hulu subscription in the student plan.
6. New features
- Blend, September 2021
- Podcast Playlists, June 2020
- Spotify Kids, March 2020
The value of personalization and recommendations
So why is Spotify investing so much time and money in bettering its recommendation system?
The first reason is that it differentiates Spotify from its competitors such as Apple Music and YouTube music. It is widely considered that Spotify’s recommendation system is better than Apple Music’s. The truth is, in all content businesses, music content or movie assets, content is a commodity. Any song sounds more or less the same on all these platforms. And anyone with enough money can buy the licenses and build their own library of content. Therefore, Spotify needs to carve out a niche to differentiate itself from competitors—building the best recommendation system.
I will assume this is also similar for video streaming businesses but I am also aware that different video streaming services such as Netflix are investing a lot of money to provide the best video streaming quality (e.g. 4K vs. 1080p) while reducing buffer time and in-app response time. So maybe in this sense, it is slightly different than the music streaming world. But nonetheless, it is still 100% true that content is a commodity.
The second reason is that the performance of the recommendation system will directly impact user retention. A good recommendation makes users more likely to continue with the service. And the more a user uses Spotify, the more the app understands the user, and the better the recommendation will get. Also, keep in mind that any personal data that a user puts into an app raises the “switching cost” since the user will have to input that data into other apps. On that front, one common challenge for any recommendation system is the “cold start”, meaning that it is always hard to provide the right recommendation at the beginning for a new user as there isn’t enough data about this user for the system to give an effective recommendation.
In conclusion, investing in personalization and recommendations is great for both listeners and the company — a savvy business move as it helps differentiate itself from its competitors as well as effectively retains users.
What are the key metrics Spotify looks at when evaluating recommendation performance?
One valuable lesson that Spotify learned is “Define success metrics BEFORE releasing your test³”. The following are the 3 key dimensions³ when considering recommendation performance:
Reach: How many users are you reaching?
Depth: For the users you reach, what is the depth of reach?
Retention: For the users you reach, how many do you retain?
In the example of “Discover Weekly (DW)”, the key success metrics Spotify defined are:
Reach:
DW WAU / Spotify WAU
Depth:
DW Time Spent / DW WAU*
Retention:
DW week-over-week retention
*In the slide deck presented by Spotify’s Chis Johnson in 2015, the Depth metric was defined as “DW Time Spent / Spotify WAU”, but I think it makes more sense if the denominator is “DW WAU” rather than “Spotify WAU” as we are focusing on the users we already reach when analyzing Depth.
How “Discover Weekly” came to fruition
“Discover Weekly” was built out of a Hack-week project at Spotify in New York City. And here are the steps they went through to bring DW to fruition, according to Chris Johnson’s presentation in 2015.
Steps to bring DW to fruition
The whole process can be briefly summarized as:
Ideation — Validation — Prototype — A/B Testing #1 — Refinement/Iteration — A/B Testing #2 — Release
However, there are quite a few steps that are worth digging into.
Prototype
Initially, prototypes are shared with employees within the company to test them out. And it got really positive results! The results are based on the primary metrics of Reach, Depth, and Retention:
- What % of users did we REACH during the week (user that streamed at least 30 seconds from DW)?
- How deep was their engagement (what % of DW users streamed 5+ tracks)
- What % of DW users came back the following week?
A/B Testing #1 — Release AB test to 1% of users
In the first round of A/B Testing, Spotify released DW A/B test to 1% of users. A Google form was used to collect user feedback and surprisingly, more than 90% of users really liked the feature and saw the value in it.
A really smart move by Spotify during this stage was that it also leveraged the image tile on DW playlists and did some experiments around it. Eventually, it found that personalized images resulted in a 10% lift in WAU.
What happens after the launch of DW?
Apart from refining the algorithmic recommendations, Spotify has been focusing on iterating content quality and enhancing its interface by taking into consideration user feedback in the form of either as direct as a question on its customer service forum, or as subtle as a user action within the app.
5 Lessons learned in the creation of DW
Here are 5 lessons³ Spotify learned throughout the process in building a great product that user loves:
1. Be data-driven from start to finish
Throughout the development of “Discover Weekly”, Spotify did not just conduct one A/B test. It conducted several rounds of A/B testing at different stages of the development lifecycle to validate their ideas and implementation, and to make sure they are headed in the right direction.
Just like what has been mentioned above, they have learned to always define success metrics before the beginning of a test, and always be data-driven.
2. Reuse existing infrastructure in creative ways
One reason that DW was able to transform from a “Hack-week project” to an actual feature in the product is that it leverages the recommendation pipeline within Spotify so that it is doable and would not greatly impact other existing features that are closely tied to the existing infrastructure.
(Here is a cool resource and impressive work about Spotify’s recommendation models if you want to get really technical: Recommending Music on Spotify with Deep Learning)
3. Don’t scale unless you need to
There will be some really big challenges if you decide to scale 100%³. Take DW for example:
Rollout challenges:
- Create and publish 75M+ playlists every week
- Downloading and processing Facebook images for implementing personalized images for the playlists
- Language translation
Weekly refresh:
- Time-sensitive updates
- Refresh 75M+ playlists every Sunday night (Oof!)
- Timezone issues need to be taken into account
4. Users know best. In the end, A/B Test everything!
From the success of Spotify’s “Discovery Weekly”, it is undoubtedly that A/B tests played a huge role in refining a product and the key to a successful product is to understand users!
5. Empower bottom-up innovation in your org and amazing things will happen.
At the beginning of this section, I talked about how the idea of “Discover Weekly” originated from a “Hack-week” project. A Hack-week is a week where employees can their interesting and innovative ideas to work and for that week, they only focus on that project they want to do. It is important for companies to realize the importance of not only encouraging their employees to innovate but also providing them the time and resources to explore those ideas.
“Discover Weekly” has also helped indie artists promote their work
So far, we’ve only talked about how Spotify’s playlist recommendation helps one of its user segments, specifically the listener user type, find the songs they like and thus increasing Spotify’s revenue by improving user acquisition and retention. During my research for this article online, I stumbled upon a Youtube video (see the following) where an indie artist talks about how “Discover weekly” helped him promote his work as an indie artist and earn some money. By drawing examples of positive interactions with people and with the recommendation feature, he says he feels encouraged to create music as an indie artist, develop his hobby outside of his day job, and get recognized by friends or even strangers who simply love his music, all while earning some extra money on the side.
It is empowering and inspiring to know that “Discover Weekly” not only helps listeners find the music they like but also encourages indie artists to get their music pieces promoted to a broader audience. The effect that the recommendation feature entails has psychological and emotional layers, which build a strong bond between Spotify and indie artists, and this is exactly what distinguishes Spotify from its competitors such as Apple Music. In addition, this is a perfect representation of Spotify’s mission —“to unlock the potential of human creativity — by giving a million creative artists the opportunity to live off their art and billions of fans the opportunity to enjoy and be inspired by it.”
Questions that emerged during my research…to be continued…
- When I revisited the homepage after a few moments (roughly 5 min), I realized that some of the trays had been refreshed, such as “New Release from [Artist I have listened to]”. What is the refresh frequency/window for a given tray? Does it happen regularly (say every x min) or randomly or somewhere in between?
- What are the things that Spotify has taken into consideration when selecting the 1% of users to test DW in the first round of its A/B testing?
Once I find out the answers to these questions, I will either add it to this article or other future articles. :)
Thanks for reading!
References
- Mehta, N., Agashe, A., & Detroja, P. (2017). Swipe to unlock: The primer on technology and business strategy. Belle Applications, Inc.
- https://qz.com/571007/the-magic-that-makes-spotifys-discover-weekly-playlists-so-damn-good/
- https://www.slideshare.net/MrChrisJohnson/from-idea-to-execution-spotifys-discover-weekly/20-2008_2012_2015Slide_from_Dan
- https://techcrunch.com/2021/08/31/spotify-officially-launches-blend-allowing-friends-to-match-their-musical-tastes-and-make-playlists-together/
- https://newsroom.spotify.com/2021-08-31/how-spotifys-newest-personalized-experience-blend-creates-a-playlist-for-you-and-your-bestie/
- https://newsroom.spotify.com/2018-11-02/our-spotify-cheat-sheet-4-ways-to-find-your-next-favorite-song/
- https://benanne.github.io/2014/08/05/spotify-cnns.html
