Spotify’s Mathematical System For Determining Your Music Taste
HTML-код
- Опубликовано: 8 дек 2022
- Subscribe to Nebula at go.nebula.tv/hai
Get a Half as Interesting t-shirt: standard.tv/collections/half-...
Suggest a video: halfasinteresting.com/suggest
Follow Sam from Half as Interesting on Instagram: / sam.from.wendover
Follow Half as Interesting on Twitter: / halfinteresting
Discuss this video on Reddit: / halfasinteres. .
Video written by Ben Doyle
Check out our other channels: / wendoverproductions
/ jetlagthegame
spotify never quite gets my music taste right, i just kinda wander along until i find a band/artist i like and then listen to them obsessively
Certified youtube music user badge, that's me too
Spotify is fucked for me. I exclusively listen to metal/heavy metal. Guess what they recommend me? Jazz.
@@Scrod117 They're... not that wrong, I'd say.
@@joaovmlsilva3509 RUclips Music is a lot better at recommending in my opinion as well
Who?
I find that spotify isnt that great for discovering new music: it kinda just recommends stuff close to what it knows you like and after a while, listening to similar music gets a bit boring imo. However, finding some new songs on youtube and then playing their radio on spotify works very well, so im pretty satisfied overall.
It's the exact opposite for me, Spotify has suggested me everything from Brazilian funk to Vaporwave. And I always look forward to what my discover weekly will have. It's interesting how different people can have wildly different experiences with the same algorithm.
I switched back to Pandora for that reason. Their algorithm recommends more unique stuff so I don't just keep playing my same playlists over and over again.
Yes, I mean, I like synthwave but I also want something different, Spotify keeps insisting in synthwave each weak :(
This.
I have the same experience. When I initially started listening to Spotify, I searched for songs that could be categorized as jazz, rhythm & blues, and similar. Nearly all the recommendations fall into the same categories, even though I do enjoy other types of music, too.
I really wish I could just see the genres directly. It baffles me that Spotify doesn't display the rest of the metadata.
It would be so cool to see what Spotify thinks I like so I can actually search for that type of music - without having to wait for Wrapped once a year.
Makes me miss Pandora but their bitrate sucks worse that RUclips
if you search spotify for 'picked just for you' it will bring up hundreds of custom playlists of different genres.
Genre tag is the one mp3 feature I need in streaming service
@@BrandonSchabes it's very bare bones though
You could make a last fm account and track all of your music data that way
@@anon5760 That's my whole point tho - every other music service shows you all of the metadata. Album and Genre are like the most basic info that all music has had since pre-internet, and Spotify goes out of their way to hide it.
I wish there was some type of voting system on the discover weekly to tell spotify if it's good or not. This week specifically I got a lot of trash in mine.
There is one. You can Say that you dont like some music on Discovery weekly. Im Not sure if it impacts though
@@KrolPawi It does. My discover weekly used to recommend a ton of gospel music to me and I hated it. So I started disliking every single one of them and after a few weeks it completely removed them from my recommendations.
Pandora has you covered.
I've gotten much better recommendations since I started clicking the heart to tell Spotify which songs I like best. I'm careful not to "Like" too many songs by any single artist. My personal playlists are really eclectic, too - like Janet Jackson, Megadeth, Jackson Browne, Five Finger Death Punch, The Mamas & The Pappas, Iron Maiden, Alanis Morrisette, Slash, Little Big Town, Jimi Hendrix, Damn Yankees, Billy Joel, Ozzy Osborne and a bunch of one-hit wonders on a single 6-hour playlist list . I've been getting really mix lists lately:-)
@@KrolPawi is that the “hide song” button?
According to Spotify my favourite music is 'speedrun', which I'm not sure whether to be happy about
"I'm fast at sex." vibes
Doesn't matter if you like it...
Aslong as you do it fast!
"Speedrun"? Like tha Hamsterdance Song?
We’re finally landing?
Why do I hear boss music?
0:40 oh yeah, Ed Sheeran is a niche indie musician
I’ve never had much luck with the autogenerated playlists, but I think their “similar artists” or whatever section is usually super spot on even for really obscure stuff, and when they automatically play songs that aren’t in the playlist you are listening to, it’s also usually really good.
Lots of people here in the comments are saying it’s only good for chart toppers but maybe you guys just listen to too much of the chart toppers because I’ve found tons of great music that was exactly what I want and from groups with like 10000 monthly listeners or even way less
Factual. Only gripe is that I listen to the song radio, and given enough time my radios end up consisting of songs I've already liked
Fuuullly agree here. I think the people complaining are the ones who think they’re quirky and “indie” because they listen to Billie Eilish and Tame Impala.
The “similar artists” section can get a bit weird. There’s a synthpop group I like called Ashbury Heights, and the “similar artists” section on their page recommends a lot of electro-industrial artists, like Imperative Reaction and Assemblage 23. Fortunately, I like both styles of music, but I find it weird that Spotify puts such distinct styles of music together.
Yeah it's true for tbe song it plays outside the playlisy when it doesn't recommend somgs you already like. I discovered what basically was my favourite of last year (fluorescent adolescent) through autoplay and its even a genre i dont listen to much
I agree! I think the "similar artists" section is the best when you explore a new genre and you only know one or two names, you can quickly discover an entire scene with some majors tracks!
As an academic working on machine learning "these rectangles make the song smaller and these rectangles are borderline sentient" is a) fucking hilarious and b) honestly not a bad summary 😂😂
Spotify Machine learning is a fancy name for a Fourier Transform 😂
I mean... I couldn't give a better definition for a CNN if I wanted to.
@@alexvaldes428 they use spectrograms, this whole video is basically ripped of an article titled "Recommending music on Spotify with deep learning" by Sander Dieleman
@@alexvaldes428 honestly Fourier Transform is fancier than machine learning
@@selfification Cable News Network
3:07 Blahaj my beloved 😍
SO SOOOOOOFT
i love blåhaj
🥰
I have one, they are important
I’m pretty thoroughly convinced that Spotify only bases it’s recommendations on recent additions to your listening habits. I only ever listen to my one big playlist that’s a giant conglomeration of everything I like (currently approaching 2k songs), while frequently adding any new music I come across that I enjoy. And according to my Spotify wrapped, I listened to 2035 songs from 900 artists across 107 genres this year. However, my Discover Weekly is only ever songs related to whatever I’ve added in the previous few weeks. My recommendations would be so much better if it _actually_ looked at everything I listened to.
It would also be far more computationally expensive, so unfortunately they limit themselves to your most recent activity.
Haha me too, same listening habit, challenging Myself to keep the Playlist under 2k 😂
That’s similar to me. I enjoy Indie pop, indie folk, pop, classical and rock and all it ever gives me is indie pop
I've given up on doing genre playlists because you never keep them up to date anyway but what's been going great for me is creating a new playlist each month where I just chuck everything into. At the end of the month I create a second playlist called mixtape where I pick my top 10 from that month. It's really nice going back a few years and listening to your old mixtapes every now and again.
Perhaps they think that you are on some kind of a genre streak and thus you're most likely to prefer listening to similar songs at this time. They also use less resources this way.
Once Spotify understands you like never heard of artists, man it finds things!
Haruhi is my favorite rapper
Ehh, still showing me mediocre songs I'd have liked years ago
It's garbage
In my experience, Spotify's algorithm was terrible. It almost never guessed what I like and what not.
Once I went to sleep with some rain sound album in background and as a result, it started to recommend me almost exclusively sound of rain which I don't want anymore. What is worse, I had no way to make it stop.
Maybe my glasses are rose colored, but man, Google play music's algorithm was bang on.
Always had a nice mix of stuff I liked already, with some new stuff mixed in. I never realized how good it was until it shut down and I was forced to use Spotify, Tidal, then back to Spotify. I haven't found a new band I like in 6 months. It sucks. Spotify just shows be the same garbage over and over, even if I selected "never play this".
Oh and YT music was trash from the get-go. All Google had to do was transfer all the old tech over, but they botched literally every aspect of the entire app, then were shocked when it flopped.
For that exact reason, I use different apps for different watching/listening, and I'll delete history. Anything to keep those algorithms from trying to shove what they think I like down my throat.
@@TKettle GPM's algorithm was amazing. If you liked GPM's recs maybe try Tidal.
I love that all the comments are "But Spotify's algorithm is garbage."
@@rhettorical that's because it is.
I love spotify, but I never have much belief that AI suggestions really know much about me. Some people worry about the big data being collected, but still we just seem to get weird song suggestions and ads out of it. Really cool system though, and it at least makes some good suggestions sometimes!
Yeah, sometimes it feels like that AI is so 'smart' that it turns right around into being dumb which can lead to interesting results.
I find the Discover Weekly feature absolutely fantastic. It almost always has new songs that I end up loving, often from obscure artists I've never heard of. And I'm very picky about songs. So it's also a great algorithm at avoiding only promoting popular artists.
They really suck at prediction. For Amazon, I never buy what is recommended to me, and for Spotify, the songs recommendations are always off. It is always so weird when people say it does good predictions when all my direct experience is that it is terrible at predicting what I like or my tastes. I can only conclude that this prediction algorithm only works on people with something off about them; perhaps they are less conscious so they don't pay attention and just click what is shown to them.
In this way, algorithms aren't really predicting anything; they are suggesting things and suggestible people just go along with it
Also, virtually everyone liked RUclips better before the algorithm took over. Further evidence that it doesn't work and just 'suggest' things, since we are inundated with crap and the good stuff is hard to find, we inevitably jus click something. The irony is is that you can get rid of all algorithms and LET PEOPLE DECIDE
People are the best 'algorithm' anyway; but these tech idiots think it was smart to replace people's actual likes, desires, and such with AI. So dumb
@@elise3455 Can you call yourself picky when you like everything it suggests?
Spotify is really good with chart topping music suggestions but anytime I try and make a ska playlist it basically is just like "You want the entire discographies of Mighty Mighty Bosstones, Streetlight Manifesto, and Less Than Jake? Maybe we'll throw a Goldfinger or Mustard Plug in there to make you think we know what niche is"
@@lil.dogbyteeh it's actually been pretty good for me over the years, altho I've recently noticed that Soundcloud is probably even better, mostly because it's not as official so artists tend to post more songs on there
Spotify also doesn't know that certain genres just don't mix well in a playlist. I listen to a lot of musical theatre, video game music, anime music, movie soundtracks, and some random other stuff I like sprinkled in. Whenever they try to make a playlist for me it just gives me whiplash. One sequence I remember very well is When the Chips are Down - > My Castle Town - > Bubblegum Bitch - > Mixed Nuts - > Test Drive - > Shinzo Wo Sasageyo - > Aria Math
After that I just stuck to listening to entire albums and my own handcrafted playlists. Because that. Well. It sure was something.
This is why I prefer Pandora's algorithm so much more.
It's just a shame it's a bit behind the times on the actual playing music part.
Spotify is awful at everything but pop
@@ghostbirdofprey No algorithm is good. I am my own algorithm, and this attempt to predict things by tech is a massive failed experiment. But since algorithms are everywhere, they think they work, when all they do is make bad suggestions
Virtually every service was better before they started using an algorithm for 'advanced prediction.' RUclips, for instance, was better, not it sucks with its recommended.
My problem with Spotify is that what I want to listen to can vary wildly from one part of the day to the next. so I never know what I'm doing to listen to next when I let the server create a playlist
Exactly. Like I’ll wake up in the morning and want dubstep, but then in the evening, I’m playing 70s folk. Like I love all my music, but those playlists throw EVERYTHING together. So if I play it, I’ll go from head banging to Excision and then the next song is Joan Baez. Talk about vibe killer. Or the playlist will have The Rolling Stones playing and then next up is a migos song. Like my “New Music Mix” in Apple Music today had a series of three songs. Shanghai Doom (dubstep), then NBA Youngboy, then Lamb of God. Like I love all them, but dupstep to rap to metal in three songs kills any vibe lol
I like using a master playlist with everything in it so that's not a problem for me.
@@alexandergilles8583 just go to the made for you section they have different playlists for the different genres you like
I have this same problem
I can confirm that the Spotify algorithm gets a little bit confused if you put Metallica, Queen, Social Distortion, John Coltrane and Claude Debussy in the same playlist.
Somehow k-pop + American rap = Turkish rap and Indian pop. Not sure how that happened, but honestly they're not bad genres
I only have completely mixed playlists, because that's the way I like to listen to music
I love Dabussy, but you have to finish on da bach!
I can't say I blame spotify for its confusion there. But I also have Kaskade, Dave Brubeck, Explosions in the Sky, John Williams, Eagles, (bluegrass covers of) Queen, Debussy, and George Gershwin all in one playlist.
I put Taylor Swift and Chopin in the same playlist and I think it just stopped trying with me 😅
0:50 "three times better than one algorithm because of math" is the best quote I've heard from you
Love the satire, really well written. Especially the jokes about spotify getting peoples music taste right.
Right?! Spotify often just jumps genre on you. It has a REAL hard time keeping things consistent. A cat picking random songs would be better - at least then you'd get some random gems... Plus a cat involved in the process.
This comment was brought to you by AI
I actually clicked on this video in hopes to learn something about Bricks.
An insight into the inner workings of Spotifys recommendation system was also a good watch but if I had to quantify it, it was about half as interesting.
If you look up "benanne" and "recommending music on spotify with deep learning" you'll get a more in-depth presentation of how it worked.
Man, I can't begin to express how utterly gratifying it is to see this many people having my same attitude and experience with Spotify. I can remember, almost a decade back, when it wasn't like this. But it's like Spotify somehow just never acknowledges the problems with what their algorithm has become.
You can dig around on their own forums and find people reiterating these same issues with the recommendations turning into a semi-closed loop for people year after year, with the people working the forums always just giving canned responses and absolutely useless troubleshooting instructions. It's gotten to a point where people tag those posts with death threats, and things like "how dare you?" and "patronizing."
The way they respond makes it seem like they think it's some isolated thing, when it's a running problem that's been going on for around 5 years. Some people seem to have a great time with discovery, but so many more complain that Spotify completely stagnates on them, and rarely recommends them anything truly new, and what it does recommend isn't appealing.
Like, I really do wonder if the people coding it even GET the scope of the problems people have with how it has all come to work. It's either that, or they are just very corrupt and somehow benefiting from this algorithm that seems to burn-out and box-in so many users.
I have a theory. They got to have some metric that is supposed to evaluate how good their recommendations are, right?
Well, so had machine translation, they had something called BLEU. BLEU is really crude. You feed it a sentence you know one (or more) good translations to, and see roughly how well they overlap with the best one, word for word.
The reason that is crude is that there may be completely valid translations that don't look very much like the target sentence. And conversely, there may be translations that look almost like the target sentence, but are grammatical nonsense, or have a "not" in there reversing the sentence's meaning etc.
Back when BLEU was proposed, that was the best they could come up with. It makes sense, when you think about it: If you had a good way to rate translations, then you basically have a translator already! So for a long time, the way Google Translate worked was in fact very close to how BLEU worked!
My theory is that Echo Nest have some convoluted but not actually very good way of measuring the quality of recommendations, and by that measure their old ways (collaborative filtering and bag-of-online-words) are very good, and neural nets are actually not that good. That's how it was for machine translation for a long while, the old methods got better BLEU scores, even though the neural net based methods got better marks from human evaluators.
And so they've downweighted the neural net recommendations, or neglected developing them further, in favour of Echo Nest's old "big data" ways (which probably have the advantage that they are a lot cheaper to run).
A couple of years (7-8) ago the AI was much better, every single week they made me a super good playlist of which I enjoyed most of the songs. After they changed stuff it got relly weird. They put scandinavian folk music in there nearly every week (never been there, never listened to that stuff). It got better over the last 2 years or so but most of the time I am better of, finding new stuff on my own...
Lmaooo what I've never had Scandinavian folk music
For me its Taylor Swift never listen to her music but Spotify put it in my playlist every few weeks
bad data in, bad data out
I also get SO MUCH Swedish music recommended…like,I listen to Swedish bands,sure,no way around that as a metal fan,but none of their songs are in Swedish
Their recommendations generally aren’t that bad for me,but that one is weird
@@bosknight7837 if you listen to sabaton at all, they might be getting confused with Carolus Rex (Swedish Version) and Livgardet being in there
unbelievable that neurostep makes a cameo in a half as interesting video
What is neurostep
@@haroldinho9930 it’s a genre of electronic dance music. KOAN Sound is the best & most popular example of a group that makes it
I think youtube does a better job in constantly recommending new songs allowing you to discover things outside of your favorite genre that aren't just an extrapolation of the few songs you recently heard , with spotify it gets repetitive quite fast
What I'm learning from the comment section is that the algorithms only work for specific types of listening styles, and that others it doesn't strike the right balance for. Which, duh I guess, but interesting to see how people have such different experiences with the discovery algorithms. I'm on some crazy power user shit with spotify so I'm definitely priming it with exactly the kind of training data that will be useful in it finding discoveries for me, so it's interesting to see how behaviors probably make it easier or harder for the thing to do its job.
@3:44 That’s not a picture of a time frequency representation, that’s a blurry picture of a waveform (which doesn’t explicitly show frequency information). Also called spectrograms, time frequency representations show you what frequency content is occurring in at a specific moment in a signal. Wikipedia has a good article: en.wikipedia.org/wiki/Spectrogram. Also the frequencies shown on the y axis in the video (0-0.4 Hz) are too low for humans to hear; our hearing range is often quoted as 20Hz-20kHz.
Yep, it's a bad chart that gets worse the longer you look at it.
3:43 this is not a time frequency representation, it is just a weirdly colored waveform
THANK YOU! And the y axis is 0-0.4 Hz, which isn't even audible. Really poorly made graphic
Thank you Sam, I’ve discovered my new favourite subreddit and something I need to buy
I’m so interested in the Spotify frequency decomposition algorithm - I tried running some of my own songs through a spectrogram filter which essentially makes time vs frequency plot
Thanks, Sam from HAI! This is a great educational video!
I didn't actually know about the weekly recommendation playlist on Spotify till I watched this. I listened to each of the 30 songs it selected for me. Only 4 new-to-me songs I liked enough to want to hear again. Not sure if that ratio is a success in Spotify's eyes but I find Pandora to be much better at introducing me to music I really like.
That said, I could clearly see/hear the Spotify algorithms (algorhythms?) at work. Every song had familiar tones, beats, phrases, cliches, etc. Just most of it fell flat. I guess welcome to today's oversaturated music landscape 🤷
You have to train the neural network for it to recommend better songs for you next week. For me I started liking and hiding the songs I love and hate in my discover weekly. Then, it learnt what I wanted and constantly gave me songs I'll love.
Thankyou for doing this video, I have actually been wondering this exact question!
I see in the comments that people find the algorithm pretty hit and miss, but I've actually got more impressed with it the more I've used it - maybe this is a weird thing to say but it's by far my favourite algorithm that I use!
I am a big music person, I listen to lots, both in quantity and variety, and I find it really difficult to categorise my own taste. Yet Spotify seems to always judge my taste just right - it's never 100% and there's always some bad recommendations as well, but the number of artists it's led me to that I never would have discovered in a million years otherwise, that have become instant favourites, is insane.
So thanks Spotify for being one of this music nerd's best friends, and thanks HAI for telling me how it works.
I had a feeling they were using sonic analysis of some description lol...
Never been so excited for a video!
Spotify knows me too well, it's recommended me 8 different versions of the Pepsi Man theme over the last 3 months in my Discover Weekly
Spotify enhance playlist feature is pretty good. I’ll make a very particular playlist and somehow they know to match the “vibe” with their suggestions. I’d say a solid 30% of the songs they put in are keepers
Always wondered how their recommendations were so good. Interesting stuff. Thanks!
1:22 It'd be interesting to use that data to find potential dates. I'd love to meet a person who listens to the same music I do.
Exactly, I've thought about the same thing too
The playlists Spotify suggested to me after this year's Wrapped came out were 100% on point. I added almost every song to my "liked"
Is this Satire
Having studied these kind of things, number 2 actually surprised me a bit. I wouldn't have thought about song reviews & descriptions being used to feed an algorithm bc they're third party data. Smart actually
Most of the time, my recommendations from Spotify are songs that I already have in playlists and listen to a lot. I don't even get a lot of different songs from the same artists, and its even more rare for different genres or artists to come into the mix.
Pretty much all I listen to is my Discover Weekly because its SO good and I love how it puts little known artists in with popular songs
The problem with Spotify is they spend eleven billion dollars to figure out what I may like based on what I already listen to, but they no idea how my musical taste might change and evolve if I heard something radically new and different. That's because they only know what I like, but don't really know me. (And thank goodness they don't!) For that reason I find user playlists far more helpful than Spotify's Daily Mixes. If I can find a user with tastes maybe 75% similar to mine, that extra 25% might have some really cool music I've never heard before.
Their ‘Weekly Suggested’ playlist, in my experience, is either completely godawful or banger after banger. I’ve found my new favourite artist through it, which I otherwise wouldn’t have even considered listening to
I've tried a lot of "systems" for recommending me music based on a sample of what I like. It hasn't ever worked so far, but that's not to say I don't think its possible.
The current systems just seem to mostly just recommend more of the same group or same genre or just something completely random.
I think we've all met people with similar "taste" in music. Getting a computer to make good recommendations will be about finding the specific nuances you like in the music you listen to and extrapolating that into other songs...bonus points for doing so with artists/bands you haven't heard before. They're working on it.
I have been wondering about this for so long
The "Precure" genre is a thing and it's all music from the Pretty Cure series of anime. And I love it.
Friendly hint - NEVER lookup children's songs using your own account - I learned that the hard way. It completely botched my recommendations and 2-years later I STILL get Disney songs showing up every week (spoiler alert: there is not way to delete or re-learn - Spotify says just "delete your account and reenroll in a new one"). Sigh.
Check if a kid stole your password haha
You can go and delete kids videos from your video history on RUclips to make the RUclips algorithm not recommend kids videos anymore. You can't do that on Spotify? I'm clearly not a user.
@@User31129 Nope. I actually contacted Spotify to try to look for a resolution, and their answer was - delete my whole account, set up a new one, *manually* re-add all my 100's of custom playlist songs (because there is no import/export function), then play those songs on repeat for a few weeks, and don't forget to upgrade with their monthly "Spotify Kids" subscription for each kid so my 2 and 3 year old's can have their own apps and accounts. I guess so they can start tracking and selling their habits from birth? SMH
I don’t think their algorithm is so much about the playlists. I’ve noticed that the « liked » library, where all the musics you’ve hearted before end up is the real gold mine.
Try to put it on shuffle, skip anything you don’t like when that song comes up, after a while you’ll notice that you enjoy the majority of what is being played to you.
I explain it kinda bad but I think you get the idea
"niche indie productions" *cuts to Ed Sheeran*... I see what you did there
beautiful haha
It really works!
Spotify is really damn good at recommending me songs i dont like at all.
I have a pretty broad taste and yeah, can confirm it manages to zero in on the stuff I don't like.
I've used RUclips Music instead, and while it certainly isn't perfect, it is scarily accurate at creating playlists that fit my music taste. Although, I suspect that YT Music works differently, since it has access to the myriad data that Google would have on its users, especially since I opt in to all of their data collecting. I'm pretty sure it uses the audio data that it gathers from the mic on my phone to figure out my musical tastes, since I still listen to the radio in my car a lot.
The algorithm works well the more feedback you give it. It also weirdly seems to get worse the less I interact with it. So if I listen to all the songs every week in discover weekly and dislike the songs I hate and like the songs I like, by the 5th week, every playlist is straight fire. I just don't actively listen to 30 new songs a week, so keeping up with the playlist to maintain quality is hard..
An Idea (for a video or something): how the categorisation of all the names work, like how do they come up with 6000 names?
I can debunk this instantly. Spotify determines your musical taste by recreating your exact playlists over and over, then summarizing your favorite songs within each into its own playlist.
The Spotify Discover Weekly playlist is always amazing how spot on it is at finding songs I add to my favorites that I have never heard before.
Sonic profile reminds me of the music genome project (which became pandora). Pandoras recs always felt spot on for me
The pronunciation of timbre at 4:09 should be in the next mistakes video - it’s supposed to be “tamber”, not “timber”
(The unexpected pronunciation of the word is well-known in the music world as something that can instantly tell you who has an extensive background in music and who doesn’t)
Lumberjack: Timber!
Musician: Actually, it's pronounced-
*Gets crushed by tree
@@possumpatrol45 HOG RIDER
oh dang well spotted. but isnt it pronounced more like tom-bruh, at least thats how I say it
1:12 ... what a weird way to wear headphones.
I call it... *backwards*
The best feature of Spotify IMO is building your own playlists. I add songs I like to a playlist for awhile, then start a new one. Fun to revisit old playlists. I rarely listen to their suggested songs.
You can do that with RUclips with a much wider selection.
@@ender7278 You could even do that with WinAmp... is there anything that doesn't allow playlists?
I don’t have much luck with playlists tbh (’nor does it give me new stuff all that often that I haven’t already heard of--), but it does get quite better usually after the playlist ends and I’m given more songs.
I think they have a 4th algorithm too, what do they really really want to push to you to listen, as I get the same random junk artists shoved at me over and over again regardless of what I listen to.
Almost all their suggestions are for indie artists and nothing mainstream, presumably because they either make more money due to lopsided leverage in the deal negotiation, or have been paid to promote them
It would be very cool if spotify could introduce an algorithm modifier to give you more randomness or reach.
Whenever I fully listen to an album the next 3-5 songs are usually pretty amazing and generally find 1-3 new songs as well
Ooh, well timed! I'm nearly done with an album and was deciding song titles, now I might do so with the algorithm in mind.
0:23 - lol you mentioned bebop like it was some weird obscure genre that like 6 people have heard of
I love when Spotify wrap comes around and have not one single song on the top 100 billboard charts
I'm not a regular Spotify user but I remember one time having it play one song and it followed it up with another song I didn't know but instantly loved that was eerily similar in style. And then there's that time it followed up Istanbul (Not Constantinople) with children's music which greatly insulted my dad but at the same time it wasn't exactly wrong.
A few years go I listened to "Locomotive Breath" by Jethro Tull, and my in next Discover Weekly half the songs were by Jethro Tull and the other half were Ian Anderson. I like rock flute as much as the next guy, but that week the algorithm was crap.
Sam is just mad that his entire Spotify Wrapped was T Swift and Abba because Ben and Adam listened to them.
My experience with Amazon's music service is they try very hard to get me to listen to what they want me to listen to. As an example, I listened to a lot of Homeboy Sandman and never did he pop up on my soundtrack except when he did an album with a well known indie artist (Aesop Rock?). But I listened to a little bit of Eminem and Amazon spent a whole day playing nothing but music from him or featuring him. Combine that with the app crashing frequently and I switched to another major service.
Once I was messing around and created a playlist called "Vintage Collection of Harmonica Rock Songs". Before I even added anything it suggested a bunch of songs that fit perfectly, but as soon as I started adding them it started going based off genre not title and stopped suggesting songs with harmonicas.
Yes. Trying to make a folk a cappella playlist for instance, it won't stay vocal for long if you keep adding the suggestions.
ANYONE NOTICE pretty much all soundtrack music is lumped together in one category. whereas hanz zimmer will make very different styles of music depending on what movie he's writing it for, sub categories of soundtrack music seem to be missing.
It hasn't worked for me. I like droney, ambient music. I like Stars of the Lid a lot, but what differentiates it for me from other droney artists like Tim Hecker (I like Tim Hecker too tho) is kind of hard to put into words, particularly since it can be very dependent on mood. Spotify recommends me some ambient stuff, but it rarely interests me, because it can't understand my very specific tastes. I can't explain it either to be fair, it's just I know when I like something. Which is why I prefer more organic methods of finding music
The problem is they start feeding you the same type of music and then it creates a closed loop where they think that's what you like. Overlooking the other music your into. It's like the "step sister" problem in 🎉movies🎉
I find Spotify's recommendations amazing. I have wondered for a while if it's built on clustering users or analysing the music; it turns out both. Thanks for the great video!
The joke at the end hit hard. Prime comes with some music, and I found one playlist when checking it out that I play over and over and over again on repeat...when I need some quiet music to block out environmental sounds so I can concentrate on reading a book.
There's an uncanny valley like effect with music taste. I love AC/DC and Spotify will throw me artists that sound similar yet are absolutely awful. Sometimes it's suggestions are okay but I tend to find what I like myself.
Your problem is that because you are used to AC/DC you are able to ignore that they are absolutely awful aswell.
@@herbertunkraut that's untrue, because AC/DC IS actually the best music that was ever created
Such a complex system yet Spotify still doesn't manage to shuffle songs
You can't even sort your playlists on web player.
Spotify had two REALLY smart machine learning interns, Sander Dieleman and Aäron van den Oord. If you follow machine learning you probably have heard those names. But back then no one had heard of them. They developed the CNN which extracts the "latent representation" of a track - like putting each song on a map, only instead of a normal map which has two dimension, this map has 50 dimensions. If you have trouble visualizing 50 dimensions, here's a tip from Geoffrey Hinton: picture 3 dimensions and say FIFTY DIMENSIONS loudly, and forget about it.
Half as Interesting says they used neural network on the waveform image. That's not quite right. They turn the sound (3 seconds at a time) into a kind of picture called a "log compressed mel spectrogram", and fed that to the CNN.
Thing is, both those interns got snapped up by DeepMind, not Spotify. And I think the Echo Nest folks are a bit old school and have a bit of a not-invented-here thing, so they still rely too much on collaborative filtering and their own "bag of internet words" model despite it not being as good as the interns' feature vectors. So if it feels like Discover Weekly isn't as good as it used to, that may be it.
I rarely use discover weekly now, I like to use genre themed playlists on spotify to listen to genres I would like to explore. I listen to daily mixes if I want to listen to something familiar and I can't decide what to pick. Too bad it's only 6 playlists.
Spotify is like Netflix for music. (sea shanties according to meme makers)
And still RUclips does a better job suggesting music that I like to me, than Spotify does
Pandora has been pretty good lately, but so has RUclips. I've been on a kick of melodic instrumental metal, leaning towards the djent end of things, and the recommendations have been awesome. I built up a playlist that runs a few days without repeat, which makes work go so much faster.
Ah yes djent, the metal sub-genre metal fans cannot decide whether it does or does not exit.
I’ve been listening to hard rock and metal for most of my life but haven’t really been trying to expand what I listen to until fairly recently and Spotify definitely hasn’t been a help.
The best way I’ve found is to ask people bc word of mouth tends to always be better than algorithm.
Do you have any particular recommendations for metal bands that you’ve enjoyed recently? I’m not picky over specific sub-genre as I’m still trying to figure out what I like. But I’m quite tired of Spotify continuously recommending me the same shit & never seem to find anything new.
So any recommendations would be greatly appreciated!
@@bonebrokebuddy5248I built out a playlist on here, mostly Jacob Lizotte, Balázs Varga, Sofian Sounds Core, Post Moral, Metal Guitar Stuff etc. All instrumental. Charlie Parra del Riego does some great metal guitar covers of stuff, and then I've got some electronic stuff from Made by John too. It's pretty much just stuff to help push through the workday. I'm not a fan of screaming or growling lyrics, but love the tonal range a lot of these songs use, so these instrumentals have been great.
spotify’s algorithm is great for how new this technology is, especially with their new playlist enhance feature i discover so many new songs every week
You need a Twice as Interesting where you compare things. Like, what makes Spotify's different from Pandora's.
+
Short video: Spotify's is bad and pushes you towards popular stuff, Pandora's is great but strange.
this is the better than any spotify ad I have ever seen. it actually made me want to try that thing.
I wouldn't. It's not actually any good at recommending music. I guess unless your taste is 'whatever', then it isn't bad at recommending popular stuff to you.
Discover Weekly is sometimes full of gems and sometimes full of a bunch of songs in languages I don't speak. Some of those are also gems.
Yay a video written by Ben
Spotify’s algorithm turned me from a hyperpop listener into a harsh noise and hardcore punk listener and i will never forgive it
Blåhaj!! I love how that cute shork is slowly creeping into the mainstream 🦈🦈
the third algorithms ironically results in music sounding more similar-ish because that way it will attract more new listeners. Like how the Phonk genre on Spotify is basically just drift phonk, even though theres other kinds of phonk music, but those dont sound like that viral tiktok hit and dont get recommended as much
Remember the old random youtube mixes/playlists? discovered hundreds of songs back in the day.
Does Spotify’s algorithm also not work for anyone else?
Thank you for justifying my reasoning as a DJ for NOT using spotify. I want my music algorithm to recommend new music that's different from what I like in addition to the same, even if it's within the same genre or genre family. Genre families are extremely relevant for electronic music, where many listeners can't even decide what is and isn't a specific genre, and many artists are defined as multiple genres or totally different genres depending on who you ask. An algorithm that does what you say it does, will only function how I need it to, if I've already listened to some of the tracks in every genre of electronic music that currently exists, or might ever exist, that I've already listened to. Yes, the other algorithm that picks similar tracks others have listened to, might account for this issue, but I don't think it would do so sufficiently. Algorithms like this tend to work similarly to selective breeding, ultimately curating users into their own little walled garden, if all they do to listen to music is go through spotify algorithmic recommendations.
The implications of this, to me, are honestly mind boggling and scary. This is how people who listen to Alex Jones and Joe Rogan get radicalized rapidly, or people who used to only listen to music previously, are suddenly getting pulled in by conservative talk radio. A conservative individual would have the same concerns over it pushing ultra-liberal content as well, relative to what I just expressed. I am not saying both are inherently the same, just that any tendency towards radicalization is ultimately potentially dangerous, even if the particular nuances of some radical change may ultimately be beneficial for all, since just as many could result in the deaths and oppression of billions of people. I don't think this is a hyperbolic risk either, given the meteoric rise of Joe Rogan's popularity once he entered into his contract with Spotify, and the very obvious results this has had on politics both in the US and world wide.
When it was just about music, it was fine, people can choose where and how they get their music and there are still many other good options (DI.FM being what I use for my streaming input to my choices). Once they started hosting political talk radio podcasts though is when it stepped into legitimately dangerous territory. I think if they want to host such media without the same risk of radicalization, short of banning specific people, would be to simply eliminate anything explicitly political from the algorithm entirely, and require that people search for it if they want to view it. Yes this would effectively ban them from getting recommended to new listeners, which is entirely the point. People with similar music interests should not be getting lumped together based on their political interests, and as far as I know this algorithm doesn't seem to filter for that, based on what I've read online, and what myself and others have personally experienced when doing testing on the platform manipulating the algorithm in various ways intentionally.
Anyway that's my essay on why the Spotify algorithm is terrible. Others have also mentioned that it honestly sucks for over-recommending irrelevant shit that people only listened to once and don't want the algorithm utilizing for their recommendations. I already have to curate my youtube algorithm by deleting random videos that i'm forced to watch on platform due to age restrictions and creator limitations, but which are irrelevant to my typical viewing preferences, just to prevent it from suddenly rapidly changing what it recommends me continuously, to the detriment of my ability to find any interesting and meaningful content. My subscription list is already so long that I don't even have time for new content at this point anyways, that if I could simply force it to use that data, plus the non-subscribed channels that I regularly seek out, maybe manually added to the list, that'd be fucking great. More user control over the algorithm would just be great in general, which I think is an extremely overlooked aspect of being online today
I wish Spotify would give us a TRUE random shuffle option instead of an algorithm based on things we liked. My likes songs sections contains everything and randomising it on shuffle is like just listening to a playlist with the same vibe rather than going from an operatic number to a heavy metal ballad.
At the very least give us a slider/switch option of true shuffle and a similar shuffle
I actually started appreciating the music in my discover weekly more since I know there are these complex algorithms behind it.
Next you should make a video on the shuffle algorithm of Spotify. From what I’ve gathered it’s not “true” random but I’m not sure the exact algorithm it uses and I’m sure many would be curious. Great video!
there's ton of info out there about the non random shuffle.
Glad I download my music from RUclips using either free portable software or a free RUclips mp3 conversion website instead of waisting my time, money and privacy with music subscription services that make a profile of me for data mining and advertising purposes!
> spends time downloading music from youtube instead of streaming it
iM gLaD iM nOt WaStInG mY tImE
very informative video. but the graph you showed at the beginning of "sonic profiles" does not look like a time-frequency graph of some piece of music, but just like a coloured waveform of something that was probably a vocal track or something. not that it makes a difference to your video's message but just in case anyone wondered how that graph would represent a song's frequencies like that
blackened deathcore was one of my most listened to genres this year :D
The problem with these NLP classification algorithms (Latent Dirichlet Allocation, Latent Semantic Indexing) is that they are intrinsically superficial ; they only match patterns by proximity, not by meaning. Neural networks are still not much better because the classification is still morphological ; it only accidentally relates to the intent (mood) of the music. Remember that Spotify isn't interested in helping you, the user, get salient recommendations, but to push content that from other artists that seem to resemble what you've listened to. It's a commercial exercise.