humans vs. algorithms: the RIYL listening exploration cycle

The "Recommended Songs" thing on Spotify playlists is based on playlist co-occurrence. I also have an internal variant that is based on listening patterns instead of playlist patterns, so here's what THAT one thought for these inputs:

https://open.spotify.com/user/glennpmcdonald/playlist/3X179t4DX6atJImPJWxL7x

is that true for "Playlist Radio" as well?

Playlist radio is also based on listening (as opposed to playlist appearances), but for "radio" purposes it tries to pick the most popular artists among the ones that are somewhat collectively similar to the seeds. The generator I used (for the link I posted above) starts from the same data, but is more biased towards the most similar artists over the most popular, and goes for more artist diversity...

(There are lots of algorithms...)

Great thread idea.

hopefully that also illustrates how loose the "playlist" can be. to me all those songs make sense together

they seem pretty cohesive, the go-betweens one doesn't have the same 80s pop sheen of the others but they all share a similar emotional colour. Don't know if mine all quite catch that but I've tried to pick some that might be less familiar on your shores:

Aztec Camera - Good Morning Britain
It's Immaterial - Driving Away from Home (Jim's Tune)
The Lotus Eaters - The First Picture of You
Icicle Works - Little Girl Lost
The Colourfield - Thinking of You
Strawberry Switchblade - Since Yesterday

This will be good.

I made a playlist too:

Class McCombs, “Don’t Vote”
Tim Hecker, “The Piano Drop”
EMA, “California”
Current 93, “The Blood Bells Chime”
Jackson Browne, “These Days”

Glenn, have you (or anyone else) shared the evaluation metric you use (or your evaluation methodology generally)?

For most things at Spotify we start by looking at depth (session lengths for whatever the feature is) and retention (what % of people come back and use the feature repeatedly over time), and how/if use of the thing correlates with overall Spotify usage and retention. For some things you can also look at save/heart/add-to-playlist rates to see if we're getting people's attention. But all of these dynamics can be very different for different features, or even slight variations on how the "same" feature is presented. We also do a lot of qualitative evaluation by looking at lots and lots of examples ourselves. But these are mostly aggregate evaluations, which don't necessarily tell you anything about how well any individual case is going to go...

Fun thread idea!

Recommended for Karl:

Translator - Everywhere That I'm Not
The Rave-Ups - Better World
Romeo Void - Just Too Easy
Michael Penn - Walter Reed
Voice of the Beehive - Monsters and Angels

Recommended for Allen:

Jack Hayter - Blind Man's Fog
Ulver - Southern Gothic
Tindersticks - Dying Slowly

so the only song mentioned in this thread so far i've heard is "little lies". i'm afraid i'm going to have to concede this to the computer.

real human RIYL for etaeoe:

f.j. mcmahon "five year kansas blues"
maitreya kali "ice and snow"
widowspeak "my baby's gonna carry on"
demonlover "ate a cicada"
stargate "driving hyperreality"

i've only heard two of the so-far 10 human suggestions for karl here. by contrast i've heard about 70% of glenn's spotify playlist. for that reason alone i think the humans are winning, just because discovering new things is so much more fun.

thank you very much ledge and mick signals for your recommendations, i'm listening through them now! ledge, you're spot on about the go-betweens not having the same sheen as the others. i kept hesitating on adding it, for that reason, but got impatient and just went with it)

here are the RIYLs from our computer overlords

********CYCLE ONE RESULTS*******

Input:

Bruce Springsteen - Brilliant Disguise
B-52's - Ain't It A Shame
Fleetwood Mac - Little Lies
Go-Betweens - Love Goes On!

Algorithmic Output (via Spotify's "Recommended Songs based on the songs in this playlist")

The Church - Under the Milky Way
Prince - U Got the Look
Heart - Who Will You Run To
Cyndi Lauper - Change of Heart
Huey Lewis & the News - Jacob's Ladder
Richard Marx - Don't Mean Nothing
Crowded House - Don't Dream It's Over
Madonna - Who's That Girl
Sea Urchins - Pristine Christine
Pretenders - Don't Get Me Wrong

Eager for emil.y's list; meanwhile, the machines are giving me real bad suggestions for how to continue this:

Jimmy Webb - Feet in the Sunshine
Beach Boys - Take a Load Off Your Feet
Parliament - Agony of Defeet
Joanna Wang - Feet
Blind Richard Yates - Sore Bunion Blues

can i make the humble suggestion that we hold off on adding new playslists until we can get a good amount of human feedback for the ones already posted, and have a chance to compare those RIYLs to the computer suggestions, and discuss etc?

like MS if you're cycle number 3, emil.y can be 4, i'll take ticket number 5 but i'll hold off until we can sort of process y'know idk i mean this is karl's thread his call i guess

i do think it'd be good to hold off a bit and take turns to the extent that we can, but ultimately i'm cool with however people want to go about it.

Emil.y called her slot before me! I think it's possible that more lists will yield more contributions, by giving would-be contributors like rushomancy more potentially familiar starting points, but also happy to wait.

yeah, i think you could be right about that, too! maybe we can just play it by ear (no pun intended), try to take turns for a while and see how it goes, and if it's too much of a hassle or it doesn't really generate much good discussion anyway, then start letting the playlists fly at wll

i mean honestly i'm in rushomancy's boat, too - i am not a walking encyclopedia wikipedia of music so a lot of this is new to me too. a lot of the suggestions i got (both human and spotify) are from artists i know of but am not really that familiar with. like, i've heard romeo void before but when i try to pin down what they sound like in my head, i can't really think of it. so it'll be fun to revisit them. richard marx, though...my understanding was that he was a synonym for "really bad MOR music", but that could just be residue from the stuff i mentioned upthread about that time of the 80s having a bad reputation, musically, when i was growing up.

Shall list originators provide a spotify link so everyone can start out hearing the seed songs at least?

sure, i can do that! should i also append the spotify recommendations to the end?

The really bad MOR recommendations? I would say no, not until the humans have been completely trounced and subjugated.

hooray! haha

ok, here's my playlist: https://open.spotify.com/user/weinventyou/playlist/5YECUJC0WWfj5b71h7UYSg?si=qDHhinuqRPOuLzeGI9TXhQ

i made it collaborative, so if people want to add their own recommendations directly on the playlist, go for it! i went ahead and added the suggestions from ledge and mick signals

"things like this THAT I DON'T ALREADY KNOW" is a different task than "things like this"...

True; since neither Team Human nor Team Computer was trying for the former task, it's slightly interesting to notice that the crunching of aggregate listening data produces a more familiar suggestions list than the amateur biological brain does.

sure but isn't the whole point of RIYL to hip you to things you haven't heard already?

That's one potential mode, but hardly the only one. Karl's setup didn't precisely define the task, but he did say "imagine your niece somehow gets into talking heads and but has never heard of tom tom club or fela kuti", which more suggests the mode in which there are notionally "correct" answers and we're seeing which method comes closer to finding them...

the crunching of aggregate listening data produces a more familiar suggestions list than the amateur biological brain does.

obviously i don't know what karl has or hasn't heard but i was aiming for songs slightly off the beaten track. spotify's & glenn's playlists might be a better match but 'manic monday' or 'don't get me wrong' seem too familiar to be useful.

yeah exactly

I'd like to hear more about what the people who have put their lists up prefer so far (the human or algorithm recommendations). I'm still mulling on what to list, also I don't have premium spotify so I'm not sure if I can do playlists on there and might need to ask someone else to do that for me and pass on the recommendations.

you can make playlists with the free version, at least you can on a computer, I'm not sure about mobile

hey karl, you definitely need this in your playlist:

Prefab Sprout - Bonny

^ has that radio-friendly springsteen sheen, the subtle mac-esque backing vocals, and the bittersweet lyricism and general strummyness of the go-betweens

also it's the best. song. ever. (obv)

having said that, this one probaly has more of the 'Little Lies' flavour:

Prefab Sprout - Appetite

Glenn's algorithmic playlist had Cars and Girls which I thought was a good match.

ha i didn't see that! basically just put a bunch of prefab on it, i'm at one with the droids on this

Recommended for Karl:
Suzanne Vega - Solitude Standing

I'd like to hear more about what the people who have put their lists up prefer so far (the human or algorithm recommendations).

yeah, sorry for the delay! i'm not sure how representative my response is, because on one hand i'm familiar with waaaay less music than most people on ILM, and there are still a lot of songs that are blindingly obvious to others that are like "new songs" to me. (i have a sharp memory of riding in the car with a girl i had a huuuuuuge crush on in high school, listening to music, and her saying "you don't know who Don Henley is?" and just cracking up). with all that in mind...

the humans won CYCLE ONE, for me. as far as songs and bands i had never heard before, i really liked mick signals rave-ups "better world" and ledge's strawberry switchblade song recommendations. to be honest i struggled with the voices of some of the singers in the human RIYLs - i guess it's an obvious point, but when the playlist is based on pretty guitar-y pop songs with basic structures, the affect of the singer really comes to the forefront, and i can be picky about that. i think that's a hurdle with me and aztec camera, for example - there's just something about frame's voice that irks me.

on the Algorithm side, the church's "under the milky way" came closest to the kind of music i was thinking of - it might have won the cycle had i not heard it so many times before. but not a bad suggestion at all. but in general i found spotify's suggestions to be very hit or miss (in the case of huey lewis and richard marx, very miss). i don't fault the algorithm so much - i fed it pretty generic pop songs that with features that are shared by many, many songs, and there's a very thin line between a singer who can emote authentically and a really corny one, and not everyone agrees on where that line is.

the grand prize winner is the NickB's last minute Prefab Sprout "Bonny" suggestion, which might have a lot of you rolling your eyes because i think it's a really popular song? but for me, that's a song that i sort of recognized from classic radio (maybe?) but i hadn't heard in so long. and definitely not a band or song that comes to mind when i'm in the mood for whatever kind of music that is (sophisto-pop, i guess?)

at any rate, i found several great new songs (and artists/albums) that i'll be listening to more thoroughly in the coming months, so i count this as a great success (for me, at least). i'm curious to see how it will play out with other genres of music, and what happens when the input is more obscure in the first place.

now i'm going to listen to allen's list and see if i have anything useful i can recommend

I don't really know what to do with Allen's list, personally. If you gave me the first 4 songs, I would never come up with Jackson Browne as the 5th. The algorithm I used for the first list basically ignores it, too.

https://open.spotify.com/user/glennpmcdonald/playlist/2taYiw5x1E8eSPazwWGp9W?si=Sla_AhRzQZOLbbAQz-PYfw

ahem, ok, having put this post in the COMPLETELY WRONG thread here it is in the correct place. ish.

purple canteen - brains in my feet
bobby valentin - funky big feet
melvin van peebles - come on feet
chic - my feet keep dancing
duke ellington - hot feet
ronnie whitehead - cold feet
shakti - what need have i for this, what need have i for that, i am dancing at the feet of my lord, all is bliss, all is bliss

Thanks, rusho!

I spotified my list with your additions, which are definitely sending the auto-recommendations in a particular direction: https://open.spotify.com/user/eater9/playlist/2kiV39jt5QUIhgl5ZL5p53?si=sHhyxPZXSyCjXolxoq52xQ

To be specific, Spotify was recommending a lot of popular disco. Then I experimentally removed the Chic track, which seemed to have an outsized influence, and now its recommendations suddenly include Dipset and Obie Trice and Kurupt alongside Loudon Wainwright III and the Monkees.

right, the recommendations aren't going to actually know what those songs have in common like a human would. also, pretty much every recommendation engine is going to recommend you songs with more plays over songs with less plays for obvious reasons. most of them will also pay more attention to the songs with more plays simply because it has more information on chic than it does on the purple canteen. the obie trice recs probably just means that it's keying in on the melvin van peebles instead, which madlib has on one of his beat tapes.

the thing that drives me most nuts about recommendations is to what extent it recommends me stuff _i already know_. if i was five and i had time to tell it all the stuff i already knew it'd be great, but i am forty years older than that and i don't need it to suggest i listen to this amazing new group called "the who". humans of course will do this too, but i like it when people do it, because it gives the suggestion of a certain larger context, which machine-curated aggregated lists simply aren't capable of doing at this point.

so for me, definitely most of my music recommendations are from individual human gatekeepers and only rarely do i find something new simply based on aggregate data.

"This amazing new group called 'The Who'" is, I think, you imposing attitude on something that has none. When I work on algorithms, I'm mainly trying to use listening patterns and math to organize the music space, or to help it organize itself. Different math can do this different ways. An algorithmic list like this one:

https://open.spotify.com/user/particledetector/playlist/0AuwHC7YvKxwnxivnJw7W2?si=CiCnPaoURqasfK88UlzTwg

is not intended to surprise you, it's aspiring to quantified canonicalization. But then the same data, processed differently, can also produce this:

https://open.spotify.com/user/particledetector/playlist/1SevIVvxunwuYZXXPW03lS?si=zpI8RY1iS5WX2RTZONDwtQ

which is full of stuff you probably don't know. I find literally hundreds of new things each week using my industrial-grade internal version of Release Radar. Mixed in, of course, with things I already know, but a skip is way easier than knowing an omitted recommendation is missing...

glenn, now be honest: are you with us here or are you with the cyborgs?

see, for me the second list demonstrates the challenges of generating recommendation data based on edge cases. there might be good stuff on that list, but i'd never know, because there's so much stuff that i can tell (without listening to it, though if I couldn't tell I'd rapidly learn by listening) is bottom-feeding junk that i'm not going to waste my time on it!

the kind of recommendations for music i want is for someone to hear that i like the who's "live at leeds" and, based on that, recommend me something like winterhawk's "there and back again". machine learning, in contrast, is more likely to produce a recommendation for "mad dog" by john entwistle's ox, which, whether i've heard it or not, i don't find to be a useful suggestion.

i don't have any idea how you would even get machine learning to produce a "there and back again" suggestion based on "live at leeds", because i'm not sure that album fits in terribly well to "classic rock" listening patterns. if spotify had my listening data it might gather that i like that winterhawk record, but it would (obviously) have no idea what it actually _sounds_ like, and trying to get an understanding of it based on my listening habits would actively lead it astray - "oh hey rushomancy listened to winterhawk and neneh cherry today".

now of course you can find plenty of new things using your internal version of release radar, but most of us don't have the time or the skill set to educate the software so thoroughly as you about what we personally want to hear!

When putting together a playlist, I often find a few of the obvious-ish algo-recommendations helpful to refresh my aging memory, oh yeah, the Who, love them, great idea. But of course it'd be a way more valuable service to introduce me to music I don't already know but will dig. Even just deeper cuts from classic artists would be welcome.

Semantic theme aside, the tracks recommended by rushomancy the human (which loosely, not exactly, line up with my implicit theme) SOUND way better to me than the ones recommended by the robot.

The robot is helpful when the semantic theme is a simple familiar one, like Rain, where it's probably had abundant trite humans to learn from.

the one thing capable of overcoming !!!THE ALGORITHMS!!! it seems is the narcissism of small differences

(that isn't even really a shitpost, I think it's the root of most discomfort with "the" "algorithms," considering people tend to be fine with them when their existence is not signposted)

good post rush

Can you elaborate on what you're saying, katherine? I think I like it.

the difference between a song someone loves and a song someone despises is often individual, trivial, petty, irrational, and/or dependent upon external context -- all of which are things algorithms are bad at. that difference is also something people assign huge amounts of importance and near-moral weight to. so a lot of "we need to get away from TEH ALGORITHMS," I find, is at root "how dare someone accuse me of liking Miley Cyrus's 'We Can't Stop' because I like Charli XCX's 'Boom Clap'"

(this thread is also sort of weighted in favor of the humans by dint of having a bunch of music obsessives actively trying to beat The Algorithms, but even so)

i don't think it's narcissistic to have idiosyncratic tastes that aren't rationally explicable. i think that's human. a really effective recommendation engine should in theory acknowledge that a lot of what we like is based on context and work to create it.

at which point, we're sort of teaching machines how to control our minds, so i get a lot of the "big bad algorithms" luddism. it's manifestly true that recommendation engines are capable of being leveraged for ideological purposes. the fascists have owned youtube's recommendation engine for years. i'm not some nerd john henry trying to beat "the algorithms" to prove a point - i want them to be better and more transparent!

"the narcissism of small differences" is the name of the phenomenon, I did not invent it

if you want to argue that it isn't narcissistic take it up with freud

katherine, i think your point is predicated on the idea of a listener who exists in this world but isn’t representative of the people posting itt (as far as i can tell).

many times i’ve been in stores, or bars, and asked what music was playing, only to find out like “oh this wilco? weird i thought i hated them. this song is cool tho.” same goes with streaming, whether it’s youtube or pandora or whatever. i don’t feel personally affronted if something comes on that i don’t enjoy; i simply don’t like it and just skip the track. for me it’s not really an ethical issue, and i think most people are fine liking the music they like and acepting that other people share their tastes in some respects but also listen to shit i couldn’t be bothered to listen to.

re: “beat the machines” i mean for me that’s just a joke. i’ll happily bow down to our cyborg overlords if they can start hipping me to shit as good as my friends do, or the fine people of ilm

so yeah, that’s a concept that exists. i’m just not sure that for me at least it’s particularly relevant to this discussion, which in any case i take to be a series of experiments that might yield interesting results on both sides — not some kind of take-down of the idea of the spotify algorithm or whatever

"because there's so much stuff that i can tell (without listening to it, though if I couldn't tell I'd rapidly learn by listening) is bottom-feeding junk that i'm not going to waste my time on it"

The robots, at least, are neither dismissive nor smugly dismissive.

(There are no robots, of course. "humans vs algorithms" is really a question of whether you're willing to try music without the protective shielding of somebody assuring you that listening to it won't make you uncool. So in that sense, yes, I'm on the side of curiosity and human potential, which ironically here we've labeled "algorithms". I very much believe that "not famous yet" doesn't equal "bottom-feeding junk". I find the robots useful not because I've trained them to cater to the narrowing of my tastes, but because I've trained me to not want that.)

(Also, there are no robots.)

To add the maybe-obvious part, I also love recommendations by people. I'm not for algorithms and against people, I'm for people and algorithms, and against "vs".

i think your point is predicated on the idea of a listener who exists in this world but isn’t representative of the people posting itt (as far as i can tell).

possibly -- "fear of being revealed as uncool" is probably closer to what I meant. or rather, "fear of being revealed as basic," since most of the thinkpieces aren't panic that they'll end up hearing 2009 Eurovision semifinalists or obscure basement folk dubstep or 14-year-olds lipsyncing to 15-year-olds singing over bad chiptune covers of Deltarune, but that they'll end up hearing the contents of the next Coachella lineup.

anyway, sorry to further derail; I'd offer recommendations but no one's list so far overlaps with what I could recommend

i'm not saying The Stolen Moans are "bottom-feeding junk", I can't tell one way or another not having heard them...

OK, you know what, I listened to them. They're some weird fur supremacist group. I have no idea what to make of them and based on the 229 Youtube views for their video neither does anyone else. They probably deserve to be somebody's favorite group but they're not mine. The video was definitely more interesting than the song.

Anyway, I'm not talking about that, I'm talking about stuff like "Bohemian Rhapsody (Arranged for Piano, Strings, & Chorus by Rick Wakeman". When the only stuff I can identify on a list is stuff like this, it colors my opinion of the rest of the list.

People say I will listen to goddamn anything and it's not true. I will listen to goddamn anything as long as someone tells me they like them. Putting that extra step in there, a person "telling" a machine they like a song and then that machine telling me about it, doesn't work. That's a prejudice sure but part of harmonious human-machine interaction is acknowledging our irrational prejudices.

Anyway, I'm not talking about that, I'm talking about stuff like "Bohemian Rhapsody (Arranged for Piano, Strings, & Chorus by Rick Wakeman".

if you manage to seed a recommendation engine and get that in return, I'm honestly impressed

so, back to business then? allen, what did you think of your human recommendations?

bump

im on streaming and almost never engage with algorithms but it’s not like im avoiding them either? idk

ah nevermomd didn’t even read what this thread was

Years ago, the first song I entered in Pandora was Pink Floyd "Set the Controls for the Heart of the Sun", hoping they'd follow it with songs of mystery and mood.

The next song Pandora played was Captain & Tenille "Love Will Keep Us Together"

i just looked up the pompeii version of that song on a porn-mode browser

up next is simon and garfunkel doing "the sound of silence"

frankly at this point i'll take daryl dragon over paul simon

that's youtube, btw

Years ago, the first song I entered in Pandora was Pink Floyd "Set the Controls for the Heart of the Sun", hoping they'd follow it with songs of mystery and mood.

But doing so did help show your age.

The spotify playlist for "Set The Controls for the Heart of the Sun" is weird in that it picks songs with vaguely similar aspects, except not the aspects that make Set the Controls actually interesting

ok, in keeping with the theme of the thread, my "set the controls" playlist, off the top of my head:

chico hamilton quintet - blue sands
om - pilgrimage
oneida - people of the north
the slits - earthbeat/wedding song
blues control & laraaji - awakening day

you decide

^ will listen as soon as i get wifi in my house or i can make it to the coffee shop before work

i also asked my floyd nerd friends and one of them recommended "om pax om" from the film "sign of aquarius", maybe that one as well