Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

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

hackernoon.com

Post reply on HN
Print view View original

71–80 of 275 posts

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

#71

Post by chrissnell »

Asdfbla wrote: post #29

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.

I think this might be a huge problem for them and I’m pretty sure that they’re aware of it. The lack of variety seems to span all genres and radio stations. I’m really into Texas Country, Americana, and Red Dirt music. These are distinct sub-genres under “alt-Country” music but from the Spotify perspective, they’re all basically the same 40-50 songs. I can’t tell the difference between the Reckless Kelly, Robert Earl Keen, and James McMurtry radio stations because they all play the same set of songs that I got sick of months ago.

I don’t know if it’s a licensing issue—do they save money by keeping song variety down?—or is it an algorithm problem? I wish they would fix it because I’m about to bail for some other service if they can’t.

Pandora is the gold standard for good variety and new artist discovery as far as I’m concerned.

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

#72

Post by dotdi »

flohofwoe wrote: post #50
dotdi wrote: post #49

I discovered how good Spotify's Discover Weekly was just a few days ago. Basically, I liked all but 2 songs on that list. My taste in music is also rather specific, which made it even more impressive. I shall see how the following weeks fare.

It's nice the first few times, but after a while I get the impression that I'm trapped in a "Groundhog Day" loop and hear the same music over and over again. Spotify should add a slider that lets me widen or narrow the 'search area', sometimes I want to hear more similar music, sometimes I want to find more stuff at the edges where all the interesting stuff lurks.

That option would be useful indeed.

I used to browse the community forums, propose features and vote on others. Spotify, however, has been pretty unresponsive to even the most reasonable and popular proposals, sadly, I might say.

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

#73

Post by lewapkon »

Cthulhu_ wrote: post #70
iddqd wrote: post #58

Earlier quoted context omitted.

I used to listen exclusively to Discover Weekly for a while. One week a finish rap song made it to the list, and I skipped past it every time it came on. Next week there was 2 finish rap songs. Then 5. Eventually half of my discovery list was finish rap, something I have no interest of. Canceled my subscription shortly after. Their algorithms are feeding themselves. I wish they had a dislike button so I could at leas…

They do have a dislike button though, the thumbs down icon next to the play controls (on the desktop version of Spotify).

It's there only if you listen to a radio, not to a playlist.

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

#75

Post by jon-wood »

Systemic33 wrote: post #51
a254613e wrote: post #6

> Unlike Netflix, though, Spotify doesn’t have those stars with which users rate their music. Instead, Spotify’s data is implicit feedback I wish they kept it simpler with ratings, and explicit instead of implicit. The discovery weekly playlists are absolutely horrible for me to the point that I don't even bother checking them nowadays. What works better for me for discovering new music with spotify is right clicking…

Explicit and implicit rating systems will never fit everyone. But I will go as far and say that implicit rating fits a lot more than explicit, because it doesn't require the user to do extra work on top of the base goal of listing to "good" music. I feel like explicit rating scales very poorly with catalogue size. So when you have music, and as much music as Spotify has, then the work effort of explicitly rating your…

There's no reason that I can see not to combine both explicit and implicit rating. If I really love a song let me mark it as such, but also feed things that I've chosen to listen to/skip/add to my library into that algorithm (possibly with a lower weight).

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

#76

Post by anon1253 »

It's very impressive. But, I'm curious about some of the technical details. For each of these representations you basically end up with a dense vector on a song level. Which I assume you would then kNN with a user specific vector. But I've never come across a nice kNN data structure that supports high dimensional vectors in a larger than memory setting whilst supporting updates. Spotifys own Annoy is cool https://github.com/spotify/annoy, but changing or adding a song requires rebuilding the whole structure ... surely that's prohibitive at scale?

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

#77

Post by yilugurlu »

ahmetkun wrote: post #45
yilugurlu wrote: post #5

They're not anywhere near as last.fm's recommendations.

Last.fm's artist radios are the things i miss the most related to music.

Yes, exactly. I discovered tons of new artists with Last.fm radios. But Spotify is playing same songs again and again and again.

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

#78

Post by cjCamel »

Angostura wrote: post #42
cjCamel wrote: post #14

Earlier quoted context omitted.

Same. Also, anytime I get a bad recommendation I blame my daughter for wanting to listen to the Frozen soundtrack.

Time to introduce her to Moana for a change of pace.

Interestingly (and now very off topic), she hasn't even seen Frozen. She does not yet know it is a film AFAIK, and I'm in no hurry to tell her.

I believe that the concept of Frozen has been explained to her by her peer group, through the medium of hair plaits.

Three year old kids are a weird and amazing bunch.

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

#79

Post by terminalcommand »

ricardobeat wrote: post #23

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!…

Spotify's daily mixes seperate the genres/moods. I occasionaly switch to listening electronic music, but I mostly listen to rock. Spotify has one daily mix of rock, one daily mix of heavy rock and one daily mix of electronic. It doesn't mix the genres.

But discover weekly only gives recomendations for the dominant genre.

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

#80

Post by arkadiytehgraet »

If there is at least one song that I like in Discover Weekly, then I consider it to be a good mix. However, this happens once a month at best. After reading the article and all three methods for recommendations, I cannot understand how on Earth it comes up with its suggestions. Do I feed it bad data? Does it not take into account the relative "weight" of a particular track, i.e. my preference to listen to it multiple times, skip others and so on? It looks like by considering these simple points it can be improved by a large margin.
Post reply on HN