How do Spotify personalized playlists work?
Spotify builds personalized playlists using your listening history, skipped songs, likes, playlist behavior, and metadata like genre, tempo, and mood. It also compares your taste profile with others using collaborative filtering, analyzes lyrics and user-generated playlists, and factors in real-time listening patterns to recommend music you'll likely enjoy.

Why does Spotify recommend certain songs?
Spotify uses a combination of audio analysis, user feedback (likes, skips, saves), machine learning, and natural language processing to match songs to your preferences. It prioritizes songs you engage with and uses similar listener data to surface tracks - though this can lead to repetition if you don’t actively engage or explore new music.

Spotify is well known for its huge amount of personal playlists. Every day, those with an account will get updated playlists. And they aren't just your basic type of recommendations either.

There is a huge amount of music delivery:

• Personalized search results
• Personalized browse section
• Playlist suggestions & enhanced playlist feature
• Artist/song radio and autoplay features
• Personalized editorial playlists
• Special personalized playlists (Your Time Capsule, On Repeat, Repeat Rewind, etc.)
• Artist / Decade / Mood / Genre Mix playlists
• Your Daily Mix playlists
• Discover Weekly & Release Radar playlists

As music streaming platforms go, Spotify does a lot with its algorithm and data.

Each of these musical recommendations is created using a range of data:

• Demographic & geolocation profile
• Temporal patterns
• Popularity and diversity preferences
• Genre, mood, style, and era preferences
• Saved songs and albums & followed artists
• Most-played and preferred songs and artists

Your taste profile is modified as you search for and listen to new music, allowing these algorithmic playlists to reach you. Even more interesting is that if you are like Blend, those playlists also get modified by using features for each individual to listen to more music.

But let's go back to the start.

Hand-picked by Humans

Long before we relied on codes and calculations, the playlists on the music streaming platforms were created by people.

Of course, the process took longer—how could it not? But expert playlist curators searched for music to add to gorgeous playlists.

At this point, you didn't get 20 personal taste playlists; you had maybe a handful for each genre.

In the early days, music discovery was limited to hand—selected playlists on streaming services and music channels on TV and radio.

Now, a quick glance at the extensive amount of playlists and updates daily would make it impossible to imagine humans doing that. This is why there is now a special place for editorial playlists that are still created by hand.

How do the music recommendations work?

Spotify's music recommendation works like a matchmaker. It matches your taste profile to the music it has on the platform. It's a lot like how TikTok gives you videos that you can relate to.

Although there has been a lot of information about how it works, the finer details of the inner workings aren't usually easy to find. Spotify isn't the only streaming service that keeps some of its algorithm magic under wraps; Deezer, Apple Music, Amazon Music, and YouTube all have similar computations that deliver recommendations.

User tastes and algorithm

Although the playlists comprise favorite songs, albums, artists, and more, they have a defined goal. The goal of these personal music playlists is user retention. They are designed to keep you on the platform for as long as possible.

User retention generates income. The more you listen, the more labels, artists, and Spotify get paid. As you listen, Spotify collects more listening data to improve its algorithm further. Their machine learning model combines the goals of Spotify to give the desired outcome for them and the user.

Artists metadata

Each track that Spotify receives from record labels, aggregates, or indie artists will be analyzed. While not all the data is provided in all cases, where it is, here is what Spotify uses in terms of artist metadata.

• Track Title
• Release title
• Artists name
• Songwriter credits
• Featured Artists
• Producer credits
• Label
• Date of release
• Tags include genre & sub-genre, music culture, mood tags, primary language, and style.
• Instruments used in the recording
• Track typology
• Local market and hometown

The data is then processed and added to the recommendation system to be flushed into locations and to users who listen to similar artists and music.

Audio signal data

As the metadata above is put into the system, the audio itself is also analyzed. While much of the process is under wraps, everyone has access to this tool and can use it for fun. While it is not currently available for public use, Spotify will still use some version of it.

Twelve markers classify a track. These 12 points tell the machine which playlists to send the music to:

• Acousticness
• Dancability
• Duration
• Energy
• Instrumentainess
• Key
• Loudness
• Mode
• Speechiness
• Tempo
• Time
• Valence

Each of these builds a profile for the track. Spotify can then categorize the song and deliver it to users who listen to similar music. The song will also be broken down into bridges, solos, verses, and anything else meaningful.

Although the information here was available through Spotify's public API pages, those 12 audio features were used in 2013.

In 2021, Spotify released research papers that cover the computations and equations to get where it is now: Multi-Task Learning of Graph-based Inductive Representations of Music Content, which is worth a read if you are really interested in the inner workings of it.

With a shift from content to context being a big trend in music for 2024, it makes sense that a Natural Language Processing model is another piece of the puzzle.

There are three primary points applicable here:

• User-generated playlists: with playlist curations and user-generated content being more powerful, these are a useful source of insight. Using the name of the playlists and the profile of each track + the lyrics = context for the songs. If a song features a lot on a specific mood-dedicated playlist created by users, that is valuable information. It helps to affirm the assumptions made by the MLM.
• Lyrical analysis: The song Pumped-Up Kicks has upbeat music but much darker lyrics. Many years later, the same mistake is unlikely to happen again. The people, moods, feelings, places, and more mentioned in the lyrics can help with categorization.
• Web-crawled data: Media outlets, social media, blogs, and more give insight into how people describe music. Those descriptions can help further distill the context of music.

It gets more complicated, and there are more layers to it - but you can already see a broad picture coming together of just a few pieces of the magical algorithm. So your music taste and the type of playlists you listen to greatly impact what Spotify mixes you get.

What is a user taste profile?

via GIPHY

Your taste profile is just as much about what you Skip and Hide as what you listen to. Your favorite artists likely never get a skip - sending a strong signal that that genre and other similar ones are in your taste profile.

Spotify (and other music streaming services) put more weight on two things:

• Passive/implicit feedback - song on repeat, listening habits like the length of session, how long you listen to tracks, and your general musical journey.
• Active/explicit feedback - likes, library saves, kinds of playlists you make, the music you add to playlists or tap like on, repeat listens, and pre-saves.

Explicit feedback requires action from the user, and if a user takes action, that is weighted more in the recommendations than a simple listening session. In other words, just because you have listened to 12 hours of music doesn't mean you enjoyed it more or less than other music. However, taking action to click a button indicates an emotive response.

As part of Spotify's own research, they worked with ten heavy music listeners and created taste profiles based on their listening history:

Source: Giving Voice to Silent Data: Designing with Personal Music Listening History

Interesting, but mostly top-line stuff you can now find in Spotify Wrapped. The taste profiles are shared because the data validates the user's music experience - further connecting them with the platform.

Where does collaborative filtering come into it?

It isn't just a user's music listening habits that impact what they get regarding song recommendations. Other users who listen to the same artists or similar artists will have a taste profile that, in some areas, will overlap. In other areas, though, each user will have artists and music that don't overlap.

Those tracks that don't overlap will likely fit the taste profiles of users who listen to the same artists, and those tracks will get pushed through recommendations. It's important to note here, though, that some fairness issues are in play because it is easy to get the same recommendations repeatedly in different playlists.

There is a looping issue: the more you listen to a single genre, the more likely you are to get only those recommended tracks. This means that popular artists also appear more often, which is part of a larger visibility issue on streaming platforms.

Spotify combines what you listen to, what users with similar taste profiles listen to, how you and those users interact, and samples from over 700 user-generate playlists.

Issues with the Spotify algorithm

The last point, the feedback loop, goes someway to explaining one of the biggest issues with the personalized playlist machine. It has seen its fair share of complaints and recommendations due to people getting multiple playlists that are almost identical.

Listening to familiar songs can be great, but getting stuck in an endless loop of the same 20 songs can get pretty tiresome.

The main issue is that perhaps the recommendation engine is too efficient because it is so focused on delivering music it knows you like rather than throwing in some wildcards.

How can you improve Spotify recommendations?

• User engagement is one of the best ways to improve what Spotify generates for you. Although it isn't foolproof, some people actively like and skip and occasionally still get those tracks.
• So here is how to give your Spotify personalized playlists a shake-up:
• Create eclectic playlists with a wide range of tracks - go on a journey of discovery and add a mix of genres.
• Make use of the search function and look for playlists that other people have made.
• Get more explicit in the feedback, like more often, go to artist profiles, and take a more active role in your listening time.
• Check out the Discover Weekly playlist, as this is where you will most likely find music you haven't heard before (but will probably like).
• Use the Artist Explorer tool to see if there are artists that you haven't heard of before - and use the create playlist function.
• Start making time to read music blogs that cover news music
• Create Blend playlists with friends who don't have the same taste as you

Your playlists won't be completely fresh overnight, but using Spotify more actively will give you a more enjoyable, personalized experience.

For some users though, the lack of great discovery has driven them to find other platforms. And, if that sounds familiar you can transfer your Spotify playlists to other platforms that have more of a focus on music discovery like Deezer or TIDAL, and your first 600 songs are free. Check out the full list of services you can switch to here: Transfer & Sync your music library.