How Spotify Knows Your Music Tastes Better Than You

Spotify has become an audio streaming giant in large part because of its advanced personalization features that rely on Machine Learning to lock customers inside its ecosystem.

Spotify is the biggest audio streaming platform with around 350 million monthly active users. Every single of its users generates data that can be used to feed Spotify’s algorithms, thus improving the quality of the experience.

Spotify has managed to stay on top of the latest Machine Learning innovations thanks to several acquisitions. From music intelligence company Echo Nest to French audio AI startup Niland, the streaming giant increased its audio analytics capabilities over time in order to make the quality of its recommendations its main competitive advantage and therefore capture more value. Let’s look at two of Spotify’s key features that use advanced machine learning.

The ‘Discover Weekly’ playlist

This playlist was a game-changer in the audio streaming world, reaching 40 million people when it was first introduced. Each week, users got a custom-made playlist with 50 tracks, allowing them to discover new songs and artists they do not know, and that they are extremely likely to like, given their listening patterns and the songs they liked the most. This flagship feature has been built using three different types of data signals:

  • What similar users like. This is also called “collaborative filtering” and is a straightforward algorithm that analyzes the listing patterns of users that like similar songs, and makes predictions based on their preferences. Users’ preferences are typically found in playlists they created themselves. For instance, if a lot of users that make playlists featuring AC/DC and Red Hot Chilli Peppers tend to also have Foo Fighters in their playlist, chances are that if you like those two bands, you will also like Foo Fighters.
  • Web content. Spotify uses Natural Language Processing (NLP) to “label” the songs that users listen to, based on content in blog posts or web articles. For instance, if “Amy Winehouse” is often associated with the keywords “soul” and “jazz” on the web, the algorithm will be tempted to look for other songs that have similar keywords to make recommendations.
  • Technical elements of the songs. Each song file is analyzed by neural networks that produce various outputs like beats per minute, key, type of instruments, loudness… Then algorithms are used to search for songs that have similar parameters. This last type of signal ensures that even songs that do not have a lot of media coverage can get recommended.
The automated “Discover Weekly” playlist, different for each user, updated every Monday.

Spotify manages to provide a very personalized experience to its users by combining these three sources of information using advanced machine learning technologies (ie. NLP, neural nets), which are the key ingredients behind the “Discover Weekly playlist that users love”.

As a subscription platform, Spotify cares about two main elements: acquiring more users (moving them from free users to paid users) and retaining current paid users in order to capture value. The Discover Weekly playlist is great for retention: it ensures that users keep returning to the platform by creating a sense of “FOMO” in the users’ mind (because tracks disappear after a week if they are not saved by the users, and get replaced by 50 new tracks)

Home page personalization

Another crucial element of the Spotify experience is what the user sees when they open the app. It is extremely important to ensure that all the options they see on the home screen make sense, grab their attention, and cover all the potential actions that they are more likely to take. The key is to offer a compelling combination of new suggested content (what Spotify calls “exploration”) and familiar content (that they call “exploitation”). Deciding which type of content to show the user in real-time requires significant machine learning capabilities that are scalable, this is why Spotify decided to choose Google Cloud services to run technologies such as logistic regressions and boosted trees that calculate the estimated probability of a users’ first action when opening the app.

Challenges and Next steps

Given the amount of data collected about songs and user behaviors, what’s stopping Spotify from composing its own music in the future, much like today’s AIs can produce ‘original’ paintings by analyzing the work of famous artists? The barrier seems to be more ethical than technical. First, users are increasingly sensitive about the ways their data is being used, and playing with something as personal as music tastes might feel like Spotify is crossing a line. Perhaps more interestingly, Spotify is not only seen as a one-stop-shop for great audio content but also as a way to connect and stay informed about one’s favorite artists (or creators). If tomorrow’s creators are machines, will it feel the same?

We might not be there yet, but we can easily imagine an intermediate solution. Spotify has an artist version of its app called “Spotify for Artists” that enables creators to interact with their audiences, control their profile page and see their listening stats and analytics. Tomorrow, the platform could make recommendations to artists about which type of song to release next, by providing guidelines on the style, the tempo, the theme, etc… Will it be the end of creativity? Or will it be a new type of (AI-assisted) creativity?

 

Sources

Akshad Tambekar, How Spotify Uses Machine Learning Models to Recommend You The Music You Like, Great Learning. https://www.mygreatlearning.com/blog/3-machine-learning-models-spotify-uses-to-recommend-music-youll-like/#:~:text=Spotify%20uses%20a%20combination%20of,features%20is%20called%20Discover%20Weekly.&text=Spotify%20uses%20three%20forms%20of%20recommendation%20models%20to%20power%20Discover%20Weekly.

Ipshita Sen, How AI helps Spotify win in the music streaming world, Outside Insight. https://outsideinsight.com/insights/how-ai-helps-spotify-win-in-the-music-streaming-world/

Spotify Engineering, For Your Ears Only: Personalizing Spotify Home with Machine Learning. https://engineering.atspotify.com/2020/01/16/for-your-ears-only-personalizing-spotify-home-with-machine-learning/

The Amazing Ways Spotify Uses Big Data, AI And Machine Learning To Drive Business Success. https://bernardmarr.com/default.asp?contentID=1201

Previous:

Darktrace: Cybersecurity of the Future?

Next:

Legacy Furniture Bohemoth IKEA, All In On AI

10 thoughts on “How Spotify Knows Your Music Tastes Better Than You

  1. Thanks for this blog post! So interesting. I am generally not very in tune with what the new music is and my friends always seem to know. I guess I have just not mastered the algorithm because it seems to me that my friends get pushed popular/ new music way more than I do. I just get a constant stream of the Dixie Chicks. I think I might have unrealistic expectations of how much is pushed to me because I don’t really want to put the work in to tell the algorithm what I like.

  2. Thank you for the very informative post Stephane. I am curious to know if Spotify also collects user data from the browser – ie, other websites such as Youtube to improve its understanding of customer preferences?

  3. Great blog post about a great company! Thanks!
    As a long-time paying subscriber and taking into consideration the long-lasting profitability concerns of Spotify, I am wondering how Spotify can further increase the willingness to pay of its customers. Given its superior technical assets, would moving away from audio-only be an option in this respect? How transferrable are the company assets and corresponding competitive advantages?

  4. Thanks for the post Stephane! I always wondered how Spotify recommendation software works. Lately, the “Discover Weekly” and even the “Release Radar” lists are so on point for me. They recommend some pretty niche tracks and artists and I just love it. And I definitely see the ML element in it as these lists keep getting more and more curated based on how I change the music I listen to. I am not sure though about the privacy concern that you raised with consumer’s music listening patterns and preferences. I think there is a point to be made here that because the consumers are not the creators of the music, there is less association of the preference to the privacy issues in consumers’ minds. At least this is how I see it. Additionally, I actually think that if they were to address those concerns in any way, they will lose a lot on the value proposition.

  5. Thanks for a really interesting post! I thought it was interesting that Spotify built its machine learning capabilities through a series of acquisitions rather than develop in-house talent. Moreover, in a world where personalization and tailored content are becoming more valuable, the question of data ownership (as well as privacy) is becoming more prominent and relevant. In this case, Spotify is capitalizing on user data and usage patterns. Does the user have ownership rights to this and should they have it or does the user surrender those rights in the terms and condition document that we all so blindly accept before downloading the app? Lastly, I found some similarity between this and Josephine’s post about machine generated art. It would be interesting to see the education needed and adoption curves expected for machine generated music.

  6. The technical elements of songs really seems like a unique and differentiated approach over the long term. The natural progression seems to be leveraging that data to create music, helping to separate their streams from the record labels and expanding margins in a manner similar to what Netflix did with original content.

  7. This is really fascinating! Interesting to see their overarching approach for recommendations: collaborative filtering (Netflix-like) + elements of the songs (Pandora/Music Genome) + Labeling songs using NLP on web content (which is a smart way to scale!).

  8. Thanks for this interesting blog!
    As already noted, the “Release Radar” and the “Discover Weekly” features work great, I found the recommendation highly tailored!
    Do you have any thoughts on how Spotify is differentiating from similar audio streaming platforms, such as Apple Music?

  9. Definitely one of the best uses of AI / ML IMO. They are able to capture an individuals tastes while still delivering the spontaneity that comes with the traditional radio experience. How do you think they manage as audio products (podcasts, etc.) become more and more niche? Is there a way for other big tech companies to replicate the platform experience?

  10. Amazing post Steph! I agree that Spotify’s deep understanding of users has a huge potential beyond improving the app and providing a better user experience. I’m curious to see if Spotify leverages this data and loops their insights back to artists and producers to help them create better music.

Leave a comment