Spotify’s Discover Weekly: How machine learning finds new music

In my experience they got pretty close to Pandora's recommendations already, which were excellent.

last.fm for me just recommended the next big artist in the same category. Often their music styles were still obviously different and the recommendation rather put me off. People that are hooked by the complexity, details and perfectionism in "Nightfall in Middle-Earth" won't necessarily like Manowar.

I recently got one Spotify recommendation that lead me into listening through a band's full catalogue and getting tickets for their show two weeks later (Insomnium, btw). It also dug out a song that I liked in primary school but completely forgot about. They discovered that I like cheezy metal covers of 80s pop songs and add one by some obscure band to my list every now and then. I'd say my experience was often pretty accurate.

Discover Weekly works well enough for me - I do wish I could exclude devices from influencing that playlist as what I listen while gaming on PS4 is completely different from the rest of the day. It also seems to have a bias towards “big” commercial releases.

Spotify’s “intelligence” in general is a huge let down though. Radio stations are extremely limited - more like 15-song static playlists indefinitely on repeat! It annoys me to no end, same for daily mixes. I end up listening to the same songs over and over and over and over again. Maybe that’s what makes them the most money?

Last.fm was amazing at finding me new music I liked. Rdio was amazing at.. radio :D I used to go for months on the same station. I miss both a lot, and occasionally I still use last.fm or everynoise.com to generate better playlists for Spotify.

Discover Weekly is weird. For the longest time it kept recommending Finish music (roughly 30% of all recommendations), I'm from Sweden and barely ever listen from anything from Finland (or Sweden for that matter).

They have a ridiculous bias towards covers. I bet I've been recommended 50 (I wish that was an exaggeration) version of the Gladiator theme ('now we are free'). And they are all terribly bad (as in blood coming out of my ears bad). I am genuinely ashamed that someone thought that it would be a good idea to submit it to spotify - and if so how spotify could recommend it to anyone, there is no way spotify could have gotten any indication that anyone has ever liked any of those versions of that song. Similarly I get lots of game of throne covers, and now Despacito covers (never listened to that song in any variant on spotify, on purpose at least).

Thanks, that is one of my main gripes with most recommendation engines, and since you like metal, the examples really resonated with me. For me, it was Wardruna - an ambienty, folky acoustic group. Listeners often also like the "typical" viking metal bands, but damn, if I want something like Wardruna, I do not want death metal singing about the same topics, I want, say, Forndom. The same often happened in electronic music as well: if I listened to 5 relatively unknown psytrance songs I liked because they have a nice balance of melodies layered with very distinct synth sounds, then no, I don't want to liten to David Guetta next, because that is also trance-like.

In other words: yes, last.fm's recommendation was much too much based on "customers also liked", which helps in a lot of cases, but so often it horribly fails ("Customers who bought The Martian on Blu-Ray also bought this asthma medication because chance happens. Wanna try it out?").

Sometimes I wish Spotify added a bit more 'noise' to their recommendations, so to speak. If I don't listen to much music except Discover Weekly for a few weeks, I (subjectively) find that what's recommended to me more or less sounds the same after a while. Either they are afraid to insert new things that stray too far from an optimal recommendation or they forget too much of my listening history.