Funny, but I was hoping this might provide some insight into why my Spotify is so bad, even by my own tastes.
Spotify constantly queues and recommends songs to me that are so bad, I can’t even imagine how there could possibly exist any data indicating that any significant sample of listeners has ever enjoyed hearing them. Spotify has 5+ years of my listening history, and orders of magnitude more data from listeners all over the world, and yet every time I set it to recommend anything to me I just sit there pressing “skip” repeatedly until I give up.
I always blamed myself, thought I was just becoming old and curmudgeonly in my 30s. But yesterday I finally discovered the problem isn’t with me. I switched to Apple Music, and it queued up 50 songs I’d never heard, and I enjoyed almost all of them.
I can use my HomePod now, too. I’m really loving Apple Music so far, highly recommend it to anyone who thinks their Spotify account is permanently broken like mine was.
I'm not typically one to buy into conspiracy theories, but I really can't help but wonder if the Spotify algorithm favors songs that have more favorable licensing terms.
Two years ago they were already shoving Drake all over paid subscribers' playlists[0]
I've paid for spotify for more than a decade and I definitely wasn't pleased to have every single playlist recommending an artist I don't like, and it was just ridiculous how EVERY[1] playlist had his photo regardless of the genre.
I had exactly the same feeling. The daily recommendations contain 80% from albums I've either already downloaded or liked (or auto-liked, which was also a very annoying thing I had for a while). The remaining 20% is recommendations of different songs of the same artists mostly.
With New releases it is different. If I ever play, say, a song someone else sent me to listen to in a different genre, then it screws up my recommendations here for months, sometimes forever.
I'm surprised that they haven't taken a page out of Netflix's book: sign small artists that sound close enough to what's popular to their own internal record label and push them in playlists?
Don't wonder. Just be realistic. Did you ever work in an IT business environment? Remember the situations when good taste slowly morphed into business interests? Yeah, that's natural in a capitalist environment.
By no chance spotify dodged that bullet. Positivity and goodwill is fine but don't spare anyone from corruption.
Counter anecdote: I've built up playlists full of tracks from good recommendations from Spotify. The genres are relatively niche and electronic (sub-genres of house, techno, uk garage, electronica, etc). The only downside is it gets a bit too eager to recommend the most played few tracks from a given artist, but I've discovered a lot of new artists this way though.
I remember that my suggestions were pretty obscure the last time I tried the spotify algo. But actually it is pretty natural that they suggest niche or less known artists because those have shittier deals and are less expensive to stream. Also, in that position you don't want to promote the top artists for free.
> might provide some insight into why my Spotify is so bad, even by my own tastes.
That's just Spotify. The AI did a better job of figuring out what I like (138 and uplifting trance) than Spotify does (EDM). I blame it in Armin van Buuren: he is an extremely popular EDM DJ and he plays a very wide variety of EDM. If you listen to EDM chances are he has at least a few dozen sets you'd enjoy. That's the problem: he's prolific and he connects every genre of EDM to every other genre of EDM. He's a "super-connector" and I could easily see how this wrecks recommendation algorithms, all roads lead to Armin.
Basically: recommending music is not the same as recommending goods, and I think the same approach is being used.
Good point. I started to notice a lot of the Spotify recommendations sounded like half of an EDM track I might like, mixed with half of something that it should’ve never been mixed with.
Spotify also seems to think I’ll like literally anything with a house kick.
It’s an extremely difficult problem to solve though so they have my sympathy!
Nobody can really explain why they love one song and hate another, and the overlap of the Venn diagram between any two people is usually very small.
I can eat any dish at a restaurant and think it’s not very good, just ok, pretty good, great, or amazing. The same for any movie, TV show, painting, drink, book, article, ... but for a song, I either like it, love it, or hate it so much that I can’t stand it.
Music recommendation algorithms have such a narrow surface area to land on, and when they miss, they go right into a volcano.
Wow, a mention of Armin van Buren! I wonder how much discovery of his work is just happenstance. His track is the first thing on Google and Bing that gets thrust into your face when you search for "this is a test".
That's my default browser bar search term for checking that my interwebs are still wired up, and I've often thought that this was either genius or a really happy accident.
I don't think that's a bad thing per se, if the algorithm correctly identifies that you like EDM. We then come to the micro-genre discussion - how specific are you in your tastes, and how precise is the algorithm required to be?
Fans of any genre say this. If you don't like metal then it's all loud guitars and screaming. When you do like metal there is fractal complexity of subgenres all the way down.
Same here. Wonderful playlists. Constant exposure to new music. This discussion seemed to get people in the mood for the airing of grievances, so here we are.
Just a couple of months ago Spotify temporarily blocked playlist exporting and there was an uproar. And it turned out that a lot of the uproar was about those Spotify automagically created playlists. Read some sentiments, coupled with some Apple Music comparative opinions, here-
I guess it just turns out that maybe Spotify isn't for all people. Nor is Apple Music. Some discussions just draw out the edges from whatever side and it becomes the narrative.
If I remember correctly, they blocked API access for one user because that user was violating the API usage terms, namely exporting Spotify's content (their playlists) to competing services. I actually though that was pretty reasonable of them.
I wasn't aware that playlists can be copyrighted. And what's the difference between someone mechanically scraping playlists (because, quite frankly, spotify alter their playlists all the time -- I've experienced this with the Soul playlists they have), and someone hiring people via mechanical turk or getting 3 other people together to make text-based lists?
> I wasn't aware that playlists can be copyrighted.
I Am Not A Lawyer, but "a curated collection of units-of-art" seems like a very reasonable thing to protect access to, independently of ownership of the actual units-of-art themselves (overly-generic terminology is intentional, since although we're discussing playlists here, the same argument could be made for, for instance, "a particular framing/hanging of visual arts")
> And what's the difference between someone mechanically scraping playlists [...], and someone hiring people via mechanical turk or getting 3 other people together to make text-based lists?
What type of difference are you interested in?
* Effective difference in terms of output? None.
* Spotify's desire to prevent people from doing so? None (they would want to provide both - and justifiably so, IMO, since, again, the playlist _itself_ is their own (algorithm's) creation).
* Likelihood of evading prevention? The MTurk solution is less likely to get shut down, for sure. That doesn't mean that you have magically gotcha'd copyright and that corporation will cease trying to protect their IP, it just means that you have found the next step in the arms race.
> I Am Not A Lawyer, but "a curated collection of units-of-art" seems like a very reasonable thing to protect access to
I'm going to need a citation on playlists specifically being protected under copyright. I'm not sure how you can claim that what is effectually a simple plaintext list of things can be copyrighted. What original content there is being put under copyright?
As you already stated, they do not alter the music itself in the way that a rearrangement would, most media players provide fade-in/out functionality so they cannot claim that is unique, and they do not provide any supplimentary content to enhance the experience, at least not in the same way that a book published list, or internet top ten list would ordinarily provide some kind of commentary on the items.
What you're claiming here is effectually that, not even the content, but the mere titles of every single "top 50 foos" list on the internet can be copyrighted. Or that the rearranging of a book's chapters without any change to the contents, can itself fall under copyright
I'm not sure how anyone could think that this is at all a reasonable position to have? It's baffling, to be quite honest.
> I'm going to need a citation on playlists specifically being protected under copyright
And I'm not going to give you one, because, as I said, I Am Not A Lawyer (and even if I was, you're not paying me to be). But, regardless, note that I made no reference to copyright whatsoever - I said "this seems like a very reasonable thing to protect access to". I'm making no reference to law whatsoever - just to what _I_ think is reasonable for a corporation to protect. You are free to disagree.
> they do not provide any supplimentary content to enhance the experience, at least not in the same way that a book published list, or internet top ten list would ordinarily provide some kind of commentary on the items.
The collection _is_ the supplementary content. If a Spotify playlist was a quasi-random collection of tracks from across the entire catalogue, then you'd be right, but they're not - the playlists are curated and specifically chosen to fit some niche (genre, artist-relation, time period, etc.). By virtue of _being in a playlist called_ (e.g.) "1940's Smooth Jazz", the songs are demarcated as being a) relevant to the particular criteria, and b) of a high-enough "quality" (whatever that means) that they have been chosen. So, yes, the plaintext-listing of those song titles _would_ have some value, just like a list of "(only the titles of the) top ten sci-fi novels of 2019" would have some value. I would, for instance, value such a list more-highly if it came from someone whose taste in sci-fi I respect and resonate with.
> What you're claiming here is effectually that, not even the content, but the mere titles of every single "top 50 foos" list on the internet can be copyrighted
Again, I'm intentionally _not_ touching on issues of copyright as legal status - but, yes, I am absolutely suggesting that someone who has gone to the effort of curating a "top 50 foos" list, _and_ of associating it with a powerful taste-making brand like Spotify's, would be justified in perceiving that list _itself_ as a valuable and protection-worthy creation.
---
After the cut because it's less relevant to the discussion, but:
> that the rearranging of a book's chapters without any change to the contents, can itself fall under copyright
I mean...if you _don't_ see how it could be possible for someone to remix the content of an existing artwork into a conceptually-new work, by making statements _with_ the playing-with-form, then I think we're just looking at art from fundamentally different perspectives and are never going to agree. If you truly _do_ want to understand my perspective, you might try thinking about how the consumption of art is affect by the context of how it is consumed and presented, not just the sequence of bytes/soundwaves/visual-elements that compose it.
+1 for the general fantasticness. I use the app on my iphone - no problems whatsoever. They even recently updated their watch app to stream directly (without the phone being nearby). Apple music has had this feature for a while.
It feels like it's a self-fulfilling spiral of bad music. The more I listen to the recommended music, the more of the same crap keeps getting recommended to me. Even if I skip them.
I remember that Spotify actually had pretty good recommendations when I first started using it at the start of this year, because I had only listened to the songs I actually liked so far. But now I'm lucky if I find one decent song out of a hundred recommended ones.
Seems there is exactly the same sort of discussion on another HN thread about poor recommendations from YouTube, also leading into a spiralling decline to more and more rubbish
Yeah. I've commented about this in the past. I have two YouTube accounts in different languages. One has great recommendations. One is terrible.
I wonder for the one with terrible recommendations if I fell into a particular failure spiral: YouTube tried to give me some new stuff for a period, but I happened to not like any of it. The AI noted I was just clicking suggestions from my subscriptions. It eventually tried different recommendations. I didn't like those either, and clicked videos I'd seen before from my subscriptions. Eventually the AI decided I only like watching videos for channels I'm subscribed to, and now doesn't recommend me stuff outside of that. Even when I branch out into a new area, it no longer suggests related channels or anything like that.
It's not just media. Every year or so I reinstall the Swype keyboard and it works great. But after about a month or two it 'learns' from my jokey misspellings or starts learning a lot of my more obscure vocabulary and starts muddying up the very basic level of Swyping I actually want it to do.
It works great when it types out phrases as I swype and I go back and edit in whatever jokes, memes, or 10 dollar words I want to include. It starts to suck when it starts trying to include those into its predictions.
There's like a sweet spot of machine learning beyond which the machine gets more annoying than helpful.
Hilariously I just switched from Apple Music to Spotify and had the exact opposite experience! I wonder if recommendations get "stale" so you get bad ones, where as if you switch services you get recommendations based on different data sets so they're temporarily better.
In my experience Spotify's recommendations for my big anything-goes playlist are terrible, but the one's for my smaller playlist for mostly British D&B music (https://open.spotify.com/track/22Z4p2gRkDPqmw8DgVixML?si=jeb... - for a taste of the genre) can give very specific and quite good recommendations of smaller artists.
I can't speak to Spotify but YouTube Music seems to do a good job with recommendations. If I start by seeding it with a "radio station" built off an artist I enjoy a lot of the follow-up music is enjoyable and helps me to discover new bands. YouTube Music even has a free ad-supported tier, which is nice.
Hahahaha YouTube has 8 years of my Google Play Music listening history, as well as my entire music library I uploaded back in 2012 when Zune (later XBM/Groove) got killed.
I started it on "The Little Things Give You Away" by Linkin Park, and let it play. For a while it was alright, it gave me chill downtempo kinda emo music selections with some like newer stuff from Linkin Park and like Green Day.
Normally I listen to slow/soft shit from bands like 3 Days Grace, Breaking Ben, 3 Doors Down, My Chemical Romance, Good Charlotte and so on with my Linkin Park (especially the Minutes to Midnight album), so Green Day and Fall Out Boy aren't bad per se even though I like my older stuff. And I do have lots of playlists with GD + LP + FOB + other random pop punk bands.
But then it started throwing in shit like NF, Machine Gun Kelley, Eminem, Logic, then D12, Run DMC, DMX. The Smart DJ on my decade old Zune HD does a better job at building a playlist and recommending music (with the same library).
Pandora, Last.fm, GPM (after 8 years it finally started to "get" me FUCK YOU GOOGLE), Spotify, YouTube Music, Apple Music, etc. They have all done this to me, they all inevitably end up playing some sort of rap, or move to super modern pop music. I do not understand it at all beyond 'it makes them money'.
I do listen to rap music, and I do like listening to those groups and artists. I have quite a few playlists built with just that music, and aside from Eminem, they do not intersect with my emo/punk/rock playlists. If I start listening to Linkin Park, I am never going to want to go from there to early 2000s rap.
I. Don't. Give. A. Rats. Ass. if every other person on the planet listens to Linkin Park with Eminem and DMX. I don't and I tell them every time I play a song or build a playlist, or skip, or remove a recommendation what I like and don't like.
Why can my deprecated Zune appreciate that and keep my genre interests separate and accurate and a billion dollar AI algo can't?
I made the same assumption about what the link was going to be about.
My solution to terrible music algorithms this year has been to, well, give up on them entirely, and go back to the world of curated music. When Google Play Music died, and YouTube Music left an even worse experience in its place, I was done. So now people curate my music instad of machines. Whether that's in the form of letting the artist curate it for me with albums, letting myself curate my favorites with playlists, or letting a DJ curate music discovery on a streaming radio station, I find that humans do it better.
College radio stations (and their streams) are great, by the way. I can't evangelize them enough. Minimal or no ads, and some of the most interesting and creative sets I've heard in my life.
Music algorithms didn't use to (or seem to) be terrible when the initial data they fed on was already curated music collections. It allowed the algorithm to really learn what you liked then provide recommendations from people with very similar music libraries.
Now it seems that everyone is listening to songs and albums based on recommendations based on what everyone else is listening to, which is also based on recommendations, so you don't find what you really like, you find what everyone else likes and it creates a feedback loop of mediocre garbage totally unspecific to your tastes. And if you try to break out of the loop you land in a loop with people who "broke out of the loop" and you get caught in the cycle again with weirder shit. Rinse and repeat until it pigeonholes you into some inescapable nether of music you might nearly but not really be into.
I would also like to plug Green River College's radio station 89.9 KGRG FM. It has an app and I've never heard a better modern Rock/Metal station
Their recommendation engine is really bad IMO. I was hoping for a futuristic machine learning + listening based recommendation engine that could find similar songs based on what they actually sound like. Mostly they just find songs based on what other playlists your song exists in. It's really boring and not good at finding new music for me. It does OKAY and it isn't why I use Spotify, but I can't help but feel like they are missing the boat here. I remember working with Gracenote software like 15 years ago and it could fingerprint songs and find similar music at least that long ago.
e: my overall sentiment on Spotify is very positive, I just wish recommendations were better.
Pretty good! I tried it with some random cross-genre combinations of stuff I like and it found a bunch of other fairly niche bands I like. Looking forward to trying out the other recommendations. Works fine for genres other than punk & metal btw.
Blog post below on how the hate5six creator made this algorithm. Boils down to using the lyrics and community listening habits from last.fm to build a similarity graph for artists, then community detection to group similar artists.
Similar experience for me - they always shuffle me most awful songs ever - I could never rely on their algorithms to the extent that no matter the mood/kind/playlist I choose, I end up skipping 20 songs that are just horrible and I just quit Spotify.
I wonder how much better algorithm could be if they'd recognise a skip within 10 seconds of a song as "no, bad choice, never play it again".
After playing an album I select (usually soundtracks), Spotify falls into a tiny selection of 10-20 tracks that it plays on shuffle forever. It appears music discovery is incredibly poor.
...I'm also afraid to play something for my son, then the recommendations while coding will surely be polluted by that; just like if you play one or two kid's videos on youtube.
FWIW, I’ve had the exact same experience. Started a 3 month Apple Music trial and was so blown away by the number of bands I’d never heard of and absolutely loved that I am ditching Spotify. I somehow found my Spotify account recommendations either recommending songs of bands I’d already told it I liked (duh!) or new bands that I absolutely detested.
It seems to always give me the best songs first and last in the Discover Weekly and Release Radar auto generated playlists. But in the middle there are sometimes unbelievably ill suited songs for me. Maybe it’s the artificial intelligence attempting to probe the boundaries a bit.
I think it just gets stuck on old songs you played. My kids played "The Bacon Song" a few times on my spotify, you can imagine what my recommendations are like now. I wish Spotify had a way to remove "bad songs" from whatever algorithm is used.
I always thought my recommendations are that broken because I listen to two very different music directions. For now, I fixed it by finding new Music via YouTube and using artists playlists etc.. Apple Music is probably no option, though, as I'm exclusively on Android.
I've had the same problem. I like multiple genres but Spotify recommendations tend ever more toward overly harsh and angular electronica. I've tried retraining it multiple times by intentionally listening in a different direction but it never fixes the problem for long. Really frustrating.
Android does have an app for Apple Music though it's been years since I've used it. I was quite happy to pay the subscription fee but an update broke the whole app for me. It would play only a part of the song and then suddenly skip to the next one.
My music taste also diverges drastically but Apple seemed to handle it quite well IIRC.
sometimes spotify recommends songs that are so bad it makes me wonder is it some sort of test to check if i am even paying attention :)
i get good results from discover weekly for the most part though. I used to spend a lot of hours on social music sites like thisismyjam and the like, where I would stumble across a gem every 1 in 50, but with spotify it feels more like every 1 in 20 or 30, so I'm happy with that and it's also less work not switching between multiple services
other people I know have nothing good to say about discover weekly so I dunno. maybe I just have low standards!
Spotify constantly queues and recommends songs to me that are so bad, I can’t even imagine how there could possibly exist any data indicating that any significant sample of listeners has ever enjoyed hearing them. Spotify has 5+ years of my listening history, and orders of magnitude more data from listeners all over the world, and yet every time I set it to recommend anything to me I just sit there pressing “skip” repeatedly until I give up.
I always blamed myself, thought I was just becoming old and curmudgeonly in my 30s. But yesterday I finally discovered the problem isn’t with me. I switched to Apple Music, and it queued up 50 songs I’d never heard, and I enjoyed almost all of them.
I can use my HomePod now, too. I’m really loving Apple Music so far, highly recommend it to anyone who thinks their Spotify account is permanently broken like mine was.