The disconnect between music recommendations and music lovers. Part 1: The Providers.

Full disclosure: I am the founder of a music recommendation service, i will not bring that up until later. This piece is mostly focused on the general state of digital music recommendations.

On paper 2016 is great for music lovers. There are a plethora of music streaming providers, reaching far ends of both genres and people.

Yet the problem still remains, a lot of people don’t know what to listen to. In a quick survey I did of about 200 people 70% said that they listen to the same things when they would rather be listening to something new.

The problem is essentially the same for everyone. The solutions are many, perhaps too many.

The providers (Spotify, Apple Music etc) themselves solve the problem by curation (playlists), behaviour based content (weekly mix), and radio (although the latter seem to be increasingly abandoned)

Some users are happy with these solutions, but they do not solve the problem for everyone. Streaming music is the new radio, and radio is usually catered for mass markets. This leaves holes for people whose tastes and preferences to do not fall in to that category. Or what Tim Quirk said a while ago:

“Online music services need bushwhackers carving paths from one starting point to another. We’re not gatekeepers. We’re not tastemakers. We’re park rangers. “ — Tim Quirk, Head of Content, Google Play music

The solutions made by the providers suffer from a few issues.

Cant recommend songs without enough plays.

Spotify can not recommend a track that which has less than 500 plays. This makes sense from the providers perspective. Spotify and Apple, Deezer etc music have roughly 35 million tracks, (Soundcloud have 125 million but thats apples and oranges). You usually have to draw the line somewhere. Excluding tracks under a certain threshold prevents creating bad recommendations.

But a track with less than 500 plays in Spotify can be a great track, made by a good artist. In another context this track can have been played frequently in other places like clubs, festivals, podcasts, public events etc.

Example: Bjarkis I wanna go bang has roughly 220k plays on Spotify at the time of writing this (~20k a month since its release). It was however one of the biggest tracks in clubs in 2015. It was played on many podcasts, festivals all over the world. by Legends like Sven Väth, Dubfire and Dave Clarke. It has over a million plays on Soundcloud (distributed between different duplicates).

Using only your streaming providers recommendation can not give you the truth about whats popular, only whats popular in their context.

Lack in granular determination of genres

Also known as pigeonholing. 200+ genres of electronic music are categorised as “dance” or “electronic”. Spotify have mentioned they have 944 defined categories. But try to navigate to a specific genre.

Complicating this issue further, genres are in themselves fairly interchangeable. One mans house is another mans techno. If a folk artist plug in a distortion pedal then is that folk or rock? Artists change genres of music during their career or even in an album. Social tagging like last.fm is one approach, but overlapping (like if every track get tagged as “pop”) can make the recommendations mute. New and less popular artists will also get much less metadata and thus preventing them from being recommended.

Popularity bias

This issue is common in collaborative filtering. Popular items tend to be recommended more often, hence creating a loop leaving the majority of content out. There are more than 4 million tracks on Spotify that have never been played. Thats up to 15% of their entire catalog!

The connection between “similar” artists are often based on collaborative filtering. This usually means based on user behaviour.

Example: Both Spotify and Apple Music list Muse and Beck as a related artist to Radiohead. This might have been somewhat true in the 90’s but since they have diverged greatly in sound. Another related artist could (should) for instance be bands that have opened for Radiohead live, Bands like Flying Lotus, Sigur Ros, Beta Band, Handsome Boy Modeling School or even Kid Koala. Similar in sound? No. related? Yes. Muse have never collaborated or opened for Radiohead.

Looking forward

Streaming providers contain huge numbers of tracks, artists. They are faced with complex issues like provide solutions to navigate the huge catalogs for tens of millions of people. In order to do that you have to make these more generalized solutions.

However, their respective catalogs are so vast they mostly overlap. Their basic services can in some sense be regarded much like an Internet Service Provider or your cell phone carrier. And you wouldn't give much value to your ISP telling you which websites to visit, or your cell phone carrier telling you who to call.

Users who actively look for music usually find themselves reading sources like twitter, facebook, magazines, blogs, message boards etc. then search their music providers for what they find. They also of course get recommendations from friends.

but keeping up with music is tedious. This results in people missing out on great music as they end up listening to the same things.

In Part 2 we will discuss apps, API’s and different approaches to the recommendation problem.