Spotify - anyone heard of it?

yes, I was joking because it would be more work. + and saved songs works well for me. More seriously, I suppose an app could be used to copy all of the single tracks in saved songs to the starred playlist every once in awhile.

Don't even get me started on stars. Still the primary reason Spotify no longer gets my money.

let me know when that's in production and i'll likely buy one off the rack.

― ulysses, Friday, February 26, 2016 12:32 AM (12 hours ago) Bookmark Flag Post Permalink

my discover weekly's have been bad and seem to have gotten stuck in a stylistic rut

is there any way to improve this other than adding a ton of songs i like to my library?

does it learn when i skip songs after a few seconds? unlike rdio there's no way to "disklike" a song, is there?

I do wish there was a way of telling Spotify that I'm not into Lion Babe or Polica.

glenn know when that hotline bling

maybe my webcam captures the look on my face as a feature of the training songspace? I've been scolding my computer as I skip tracks just in case.

Does adding songs to your library actually affect Discover Weekly? I thought it was only based on the songs you've actually listened to.

xp more seriously, a skip would probably be weighed less since people might skip around just to get a feel for the playlist or something.

Discover Weekly is informed by your listening, not your library. So the bad news is that you can't force it to change by any quick tricks. But the good news is that if you listen to more music you like, and explore more kinds of music that aren't in that rut, it should pick up on those cues and shift along with you.

And yes, people skip for a lot of different reasons, so we're pretty conservative about what use we make of skips. They're a little better a signal in aggregate than for individual listeners, but even in aggregate there can be insidious complicating factors.

So then my method of making a massive playlist of a bunch stuff I really like and playing it on shuffle as I go to sleep but with the sound off is a good one?

Yes, if you keep doing that consistently, we should pretty quickly zero in on other music you will enjoy playing with the sound off while you sleep.

This, I assume, describes the guts of what would become Discover Weekly: http://erikbern.com/2015/01/13/scala-data-pipelines-for-music-recommendations/

maybe a better version, since it mentions the NLP and audio pieces: http://www.slideshare.net/MrChrisJohnson/from-idea-to-execution-spotifys-discover-weekly

I'm shocked that user libraries don't go into it! I feel like I used Spotify very little before Discover Weekly came out, but it still seemed dead-on from the start.

yeah, that first one was way too simple. very interesting, though, thanks!

I meant the recommendation algorithm part (not the whole thing) of the first deck was simple, i.e. it didn't seem to explicitly account for genre anywhere.

actually, nevermind. The second deck just goes into more depth.

Yeah that second one is really interesting. Seems like a pretty solid setup.

The discovery playlist function has been massively improved imo -- I actually look forward to checking them out now.

A few obvious ones this week like Sonic Youth - Cross the Breeze, Kraftwerk, Tortoise, but this is still much less obvious than the obvious stuff of yore.

"Spotify has launched a new music discovery service called Fresh Finds, which aims to shine a light on artists before they make it big."
http://www.factmag.com/2016/03/02/spotify-fresh-finds-playlists-launched/

The features launches today (March 2) with six playlists to choose from: Fire Emoji (hip-hop), Basement (electronic), Hiptronix (vocal pop), Six Strings (guitar driven), Cyclone (experimental) and the more general Fresh Finds, focusing on “breakout tracks” across the other five playlists.

maybe a better version, since it mentions the NLP and audio pieces: http://www.slideshare.net/MrChrisJohnson/from-idea-to-execution-spotifys-discover-weekly

this is interesting, and namechecks some modern stuff (e.g. they have a screenshot of the word2vec paper from NIPS 2013), but it's not obvious to me which of those approaches they're actually using (collaborative filtering, NLP like word2vec on unstructured non-spotify text, dimensionality reduction/latent spaces, and "deep learning" (lol) on the audio). in particular, i heard the audio analysis stuff had turned out to be a dead end. is there anything more technical on how they're doing this?

ok slide 46 suggests they're using all of them, or at least plan to use all of them

also hints that they want to pay attention to saves and skips but as confirmed in this thread they are not currently doing that

🤔

it looked to me like they're using collaborative filtering, and the other NLP stuff is used for describing features of songs (beyond what publishers provide) and therefore user preferences. So maybe it's collaborative filtering with implicit feedback but with a lot of cool machine learning stuff also going into that feedback?

audio analysis stuff had turned out to be a dead end

you don't say - i am stunned that analyzing key signatures and beats per minute did not result in great recommendations

Spotify is great for parties (Pointer Sisters!), but on my own, I spend more time with (listen longer) and get more out of music that I've paid for and own. I guess just the breadth of selection and access is quite paralyzing and distracting to me.

it looked like NLP to describe songs and user preferences (from playlists?) -> dimensionality reduction to describe user preference alignment -> feeding this alignment into collaborative filtering for recommendation

Spotify certainly knows I had a thing for "Some Velvet Morning" at one point in my life - I've had Nancy & Lee doing it twice and Lydia Lunch & Rowland S Howard last week. Looking forward to the Entombed, Vanilla Fudge and Slowdive versions on future Weekly playlists.

thin white rope version ftw

http://open.spotify.com/track/5yHuxE6LZuRmCPJykmCyFC

lol, PVMIC

guilty as charged

wrt "NLP" (all of which is just essentially matrix factorisation here) it looks like they're combining all the most sensible/obvious signals, as I said before it looks pretty solid. The "NLP model" and "audio model" give you similarity but not necessarily popularity, the others give you a mixture of similarity and popularity; they have a "random negatives" component so they can do discriminative training (or NCE or w/e).

And if all that sounds interesting, we're also hiring...

"Pretty solid" is an understatement; I'm pretty sure that level of engineering behind something like playlist generation is completely unparalleled. I'm happy they're so open about how it works.

in particular, i heard the audio analysis stuff had turned out to be a dead end

Then what are all those Echo Nest folks doing now?

I'm doing quite a bit of skipping as I check them out, but the Fresh Finds playlists are pretty cool. I assume they are generated in a manner similar to Discover Weekly playlists, which I also like a lot.

We use audio analysis for a lot of things, but rarely use it by itself. Spotify Running, for one obvious example, makes extensive use of audio analysis to figure out what songs might be good for running, but then also uses non-audio data to help figure out which good-for-running songs are suitable for particular running modes or particular runners.

Fresh Finds and Discover Weekly share the broad goals of exposing listeners to more music and exposing more artists to more listeners, and come from the same NY+Boston part of the company, but they use different techniques. DW is mostly a personalization thing, obviously, and these current FF lists aren't. Fresh Finds is trying to surface mostly-unknown music, so almost by definition it finds stuff that doesn't yet have the kind of collective activity that would result in it showing up in anybody's DW. But things FF elevates out of obscurity might later show up in Discover Weeklies as an indirect result.

(Audio analysis was only one of several things The Echo Nest did, and as part of Spotify we're pretty much doing more of all of the things we did as a smaller independent company.)

can you say in what sense you're using "NLP", which was alluded to in the discover weekly slidedeck briefly?

do you mean you're using dimensionality reduction techniques common in NLP on non-text data, or you're analysing press/reviews, or you're literally doing NLP on song titles and artist names?

We machine-read blog posts and news and reviews. (We also do dimensionality reduction, but that's not what the NLP mention meant.)

This is interesting, though far outside my expertise.

I haven't personally found Discover Weekly to be a useful feature yet. Seems like it could benefit from a little more...chaos? (Possibly I'm delusional that my tastes are more idiosyncratic than they are.)

x-post
You probably already do, but do you pay attention (can you?) to when a user turns a song up (volume-wise)? That's what I do when I REALLY like a track...

it looks like they only use explicit feedback like that in radio (though I don't think they use volume). perhaps this is part of the overall thinking that led to the demise of Jeff's precious stars...

I've been listening almost exclusively to the massive "everything ILX listened to in 2015" playlist, in reverse alphabetical order, so my Discover Weekly is agreeably eclectic!

glenn are you in nyc?

No, I'm in Boston. (Somerville, actually. The old Echo Nest office is now Spotify "Boston".)

This is the most detailed analysis of Spotify Fresh Finds I have seen so far in a news article:

Spotify is using 50,000 anonymous hipsters to find your next favorite song
http://qz.com/628812/spotify-is-using-an-anonymous-army-of-50000-hipsters-to-find-hot-new-songs/

the article explains the process in stages. See the diagram in the article.

For the last few years, Spotify has been gathering data from music blogs and review sites, and culling out the most talked-about new artists. It could just feature those artists in a playlist and call it a day, but there’s a hitch: many of their songs are so new that they’re not on Spotify yet.

To get around this chicken-and-egg problem, Spotify figures out who is listening to those trending artists, and then uses the hippest 50,000 or so users—the people who find about music before it is cool—as a new-music focus group.

Who are these prescient hipsters? Spotify isn’t saying, and the users don’t even know that their bleeding-edge taste is being used to create the playlist. But their favorite new songs—released within the last three weeks, with Beyoncé-level stars filtered out—are the raw material for Fresh Finds.

A group of Spotify employees then sorts the new songs into different genres (hip-hop, electronic, electronic pop, guitar-driven, and experimental), puts them into an appealing order, and voila: a weekly playlist, released every Wednesday, made up of songs that early adopters love and you might, too.

are you cool with spotify making daily playlists of our interests without asking permission?

(my answer there is 'of course'; the applicable old saw about how if you're not paying for the product you are the product makes sense when the cost of the service is so radically low... though i still wish they'd go back to the way emusic did it and hire an editorial staff)

I think that's what Apple Music does...

Who are the editors for Apple Music? Legit music journalists or celebrities who are likely having their assistants and management music select?
Serious question, i dunno.