Spotify is the largest music-streaming service and offers their customers a catalog consisting of almost all existing music. Next to numerous playlists and various radios, Spotify also offers a customised playlist: Discover Weekly. Every week, Spotify provides their users with a playlist full of new or undiscovered music. How does Spotify know the music that you will like?
Spotify uses three main technologies to recommend music to their users that they will enjoy. The first technology is Collaborative filtering. This technology analyses users’ habits such as which music they listen to, which artists they like and which playlists they follow. The technology matches users to other users with identical habits and recommends music that similar users are listening to.
Spotify also uses Natural Language Processing. This technology scans all written content concerning music on the internet. It will for instance analyse all content written about one artist, scan all the words around this artist’s name and store it in Spotify’s database. When a user listens to this artist, it will recommend a different artist that is described by similar words.
The third technology is Audio modelling. This analyses the music itself and identifies its key characteristics. It then compares this data from different music and matches music that has similar key characteristics. It will thus recommend music that has similar characteristics to their users.
By combining these three technologies, Spotify ensures that it is not only dependent on other user’s habits and written content on the internet. When music is new or unknown, no user data or written content is available. However, Spotify is still able to recommend it due to audio modelling.
Spotify is getting good at making accurate predictions about what users may like. It knows your music preferences even better than you do yourself. What do you think, does Spotify accurately predict music that you like?
Sources:
https:// www.digitaltrends.com/music/best-music-streaming-services/
http:// blog.galvanize.com/spotify-discover-weekly-data-science/
https:// www.slideshare.net/erikbern/collaborative-filtering-at-spotify-16182818/10-Supervised_collaborative_filtering_is_pretty
https:// notes.variogr.am/2012/12/11/how-music-recommendation-works-and-doesnt-work/
Hi Francis, thank you for this interesting post. I would like to start with the question you ended your blog with. I tried the discover weekly playlist several times. Unfortunately, I am not really impressed by the songs they offer in this playlist. In my opinion, music is a form of art which makes every song different. Of course, this makes it very hard to compare a song of one artist to another based on the characteristics the technologies are looking for to match.
I think Spotify can improve their recommendations provided by the collaborative filtering technique. The listening habits you talk about in your blog may be misleading for this feature. Imagine that you have created a very large playlist with a variety of genres. You do not like all these songs evenly. Spotify interprets every added song to a playlist as a like which may influence the recommendations you get negatively (Douglas 2018). I think the collaborative filtering would be more effective if users could rate every individual song. What do you think?
DOUGLAS, N., 2018-last update, If Your Discover Weekly Playlist Sucks, Try This. Available: https://lifehacker.com/if-your-discover-weekly-playlist-sucks-try-this-1823436104 [14-10-, 2018].
Hi Francis,
Thank you for your interesting post!
I also use Spotify, but I am not very enthusiastic about the Discover Weekly playlist. In my opinion, it does not accurately predict songs that I like since I often skip at least half of the songs.
I don’t think it considers the moment that you listen to certain music. For instance, the music genre that I listen to while studying differs from the genre I’m listening when I’m hanging out with friends. The discover weekly list does not distinguish between these genres which results in a list of random picked songs.
In my opinion, Spotify should first take into account the moments that users listen to music and then compose multiple lists for different genres. This will ensure that more people listen to the playlists and enhance the customer experience.
Hi all,
First of all thank you Francis for the interesting post. I would like to reply to Joran as I happened to have read an article about Spotify yesterday, which deals precisely with tailoring music types to specific moments. The article explains that Spotify is currently thinking about using motion sensors that users would wear or that would transmit data, such as temperature and heartbeat, through their smartphones. These sensors would be able to analyse the activity of the user and target music according to those moments. For example, the app would play more pumped up music as it senses that the user is working out, or conversely it would play a more gentle music during calm moments. If this came into production, the issue of music recommendation for the wrong moments would be solved.
https://www.theguardian.com/technology/2014/jan/20/spotify-sensors-heart-rate-mood-playlists-motion-tracking
Hi Francis! Thank you for this post on this up-to-date subject.
I actually used the “Discover Weekly” feature quite a bit when I started using Spotify. But I feel that over time I’ve been decreasingly prone to click on it. I think that’s mostly an issue on the volume of the playlist. What I liked it for was encountering the odd song that I didn’t know existed (and that’s also probably what they want to achieve). However, the fact that I have to listen to multiple songs, which frankly aren’t always subject to my taste, made me stop using this aspect of Spotify.
Personally, I think Spotify could apply the versioning principle and offer users various types of the recommendation playlists. For example, I actually use the number radio feature more often than not lately, purely because I can then control what kind of songs I feel like listening to at that time.
Thank you for explain this recommendation system so clearly. It still doesn’t work that well for me but I’ve heard very positive stories from others. I just wanted to point out how much good for the artists Spotify can do with this system as well. Some artists that were unknown had their songs listed in hundreds of thousands Discover Weekly playlists and in this way became famous almost overnight. Another good story to illustrate how Spotify launches careers is with the song Call on Me from Starley Hope. Her song was picked up in Denmark, got put in a playlist there and did well. Spotify than saw through their analytics how popular it was and put in a slightly more popular playlist (1 million followers) and eventually in the biggest hits (14 million followers). The song became a platinum hit because of Spotify’s analytics and recommendations.
It’s an interesting story and a good read, you can find it here: https://www.wired.com/2017/05/secret-hit-making-power-spotify-playlist/