Spotify – Your Personal DJ Committed to Curating Your Music Experience

Spotify’s use of thoughtful use of ML to curate personalized listening experiences for its users.

The ways in which we access music has evolved from analog storage modes such as vinyl, cassettes, and CDs to digital storage modes like iTunes to the current mode of streaming. Streaming gives listeners the ability to listen to any song of their choice within seconds and has had a large impact on the way music is consumed and artists are paid. With artists being paid based on the number of streams, songs on average have gotten much shorter which also mean that listeners can listen to more songs that they did before and companies can gather a lot more data about listener’s habits. Spotify has been intentional about collecting data and using it to improve its service for listeners and artists.

From a listener’s perspective, the best use of data has been Spotify’s ability to curate personalized playlists that not only take into account songs the listener already likes, but also include songs that the algorithm predicts the listener would like based on historical data. For music enthusiasts, this is a valuable feature as it allows listeners to effortlessly discover new music. Spotify’s “Go To Radio” feature also allows users to listen to music that is similar to a chosen song. This feature learns on the go by allowing listeners to “like” and “dislike” songs and updates the song queue in real time to reflect the listener’s predicted preference. Finally, Spotify’s “Wrapped” provides a yearly summary to listeners and shows everything from top 5 played songs, to most listened to artists, to how many minutes of music the listener streamed in the year. This has become a strong marketing tool for Spotify as users all over the world share their results on social media and compare their music tastes with their friends.

From an artist’s perspective, Spotify helps listeners find new artists that they ordinarily may not have comes across. Spotify has created a platform where the algorithm is able to cut through the noise and predict whether a listener will like a song, regardless of the popularity of the artist. Spotify also uses its data to let listeners know when an artist they like has released a new album or has a concert nearby. This is an additional means of targeted advertising for artists.

Over the years, Spotify has made large investments in its machine learning capabilities to improve its ability to personalize the experience for listeners. It uses ML in 3 main ways:

  • Collaborative Filtering –  this means making recommendations for a single user by using data collected from millions of Spotify’s users. With over 400 million active users, 8- million+ songs, and 4 billion+ playlists globally, there are bound to be patterns that Spotify can draw from such a large sample size.
  • Natural Language Processing (NLP) – Spotify categorizes music using NLP to determine the “vibe” of a song. If a song is deemed more upbeat, it might go in a “turn up” playlist. If the computer reads a song’s lyrics and assess them to be about love, but in a sad way, the song might go in a “heartbreak” playlist.
  • Reinforcement Learning – this means that when new content is presented to a listener (this new content was likely found using collaborative learning and /or NLP), the algorithm takes note of the listener’s action. If the listener skips the song, the algorithm will note that the prediction was weak. If the listener “likes” the song, this signals that the prediction was accurate.

When Spotify was founded in 2006, the technology it needed to become the company it is today did not really exist. The company has always had to hire talent that was at the frontier of ML technology to ensure that it had the technical capability to achieve its long-term business objectives. It is easy for an organization to get overwhelmed by the amount of data it has. Having data in itself is not the competitive edge, but knowing how to organize, use, and draw key business objectives from the data is critical. Spotify has decided that personalization is a core part of its business and all the data it receives is clearly geared towards making a platform that is at the forefront of curating a personalized music experience for its users.

Sources:

Previous:

Marriott: Data-driven Customer Experience for Decades at Scale

Next:

Asian Paints: India’s Biggest Data Science Company that Sells Paint

Student comments on Spotify – Your Personal DJ Committed to Curating Your Music Experience

  1. I’ve been a Spotify user since 2013, and the difference once they began implementing ML into their music suggestions has been night and day different. The “Discover Weekly” playlist that operates off of user preferences is a super unique feature, and they do a great job of recommending artists that are out of the mainstream, allowing them to gain larger followings.

    I’m really curious about what factors into Spotify’s recommendations– is it based simply on what similar users have liked (like Amazon’s shopping recommendations), or does it have to do with musical cues, like tempo or the key of a song?

  2. Thanks a lot for the post Sultana!

    I’m a heavy user of Spotify’s “radio” feature that you mention and it keeps me engaged with Spotify (i.e. a more loyal customer). It helps me discover new music and reduces the hassle of having to prepare a playlist or queue songs when you’re playing music at a dinner or party. I think what makes Spotify’s data analytics so powerful is the “Reinforcement Learning” that you mention: the more a user uses the app and radio, like, skip features, the better the algorithms become and the more likely the user is to keep using the app. This positive flywheel powered by data keeps me engaged with the app and increases my switching costs to other streaming providers.

  3. Thank you for the post Sultana!

    As Katelyn said, the ML algorithm really gave Spotify the edge to be basically your favorite radio station but even better, since it’s based on your own music.
    It’s also interesting how very small things in the interface have been changing to adjust to this new algorithm, before you listened to your playlist in a loop, now you start listening to one of your playlists, and without realizing, you are discovering a new song because Spotify changed the default settings (without us knowing) to allow the algorithm to continue your playlist style, and we are not even mad.

  4. I really enjoyed reading your post!

    I have been a loyal Spotify user for almost 10 years. I originally used Pandora for music streaming but I ended up shifting to Spotify because I felt that it did a better job curating my music recommendations. I didn’t realize that the algorithm was built off of a combination of three different ML features but after reading your post I now understand why the recommendations felt more curated to my own personal music interests. I had also never heard of collaborative filtering. This form of ML benefits from a larger user base because the ML capabilities improve with a larger data set (i.e. more Spotify users). For this reason, I think that Spotify’s algorithm will continue to improve as the company grows and attracts more customers.

  5. This is so fascinating – never thought about thinking of how songs have become shorter over time, but it makes perfect sense as to why! I am disappointed by their data use though – I haven’t found as much value in their recommendations, which is especially surprising given I’ve been a user for 10 years – but maybe that’s why this whole thing is so tough, music is so emotional and subjective 🙂

Leave a comment