> "if people would listen to what is more profitable to us, instead of what they enjoy listening to, we would make X more profit".
I was ranting, immaturely, to my ML educated friend recently about how terrible modern discovery algorithms are ("if ML is so powerful why does Spotify exclusively serve up the steamiest of dog sh$t?") when he explained this to me.
His response was "they aren't optimizing their discovery algorithms for your enjoyment, they optimize them for profit."
> "they aren't optimizing their discovery algorithms for your enjoyment, they optimize them for profit."
Still in 2016, I could start with a song in genre X and then let Spotify autoplay for hours staying in genre X. I discovered some new bands I liked. Then circa 2017 or 2018 something changed. After a couple of songs, Spotify would switch to genre Y, and break my flow.
I stopped paying for Spotify, and switched to Youtube, then Youtube Music. Youtube Music still does the good old thing, plays random songs staying within the genre I started with.
Maybe Spotify is optimizing for profit, but my money they lost.
If you listen to some songs off track it will ruin the recommendation engine for a long time with Spotify. E.g. child music is the most obvious one. Having a party with people choosing songs ruins it.
So in practice it is not very user friendly since you can't pause it.
I'm not sure this is a good solution; it violates multiple UX principles around the ability to undo, to see what a program is doing. It's also bundling a lot of stuff together in an unintuitive way: "Incognito" (or as Spotify seems to call it, "private") listening doesn't signal that it has anything to do with recommendations, and if you turn it on I believe it also does things like block social features.
I don't think "private mode" in applications is a good alternative most of the time to more direct controls like:
- An opt-in mode for training the algorithm.
- An ability to after leaving a phone on overnight to clear a bunch of songs out of your recommendations.
- An ability to mark a song that you like as irrelevant without blocking it.
- An ability to import and export recommendation lists and then to "pause" or "freeze" your recommendation training.
- An ability to freeze/fork/revert your recommendation profile.
- The ability to have multiple recommendation profiles.
- etc...
Going into a special mode to preemptively avoid a program from acting unpredictably and doing unpredictable things is bad UX, we should just call it bad UX. One of the big reasons that bad UX exists is because companies aren't using recommendation algorithms to benefit the user, and so they don't want users to be in control of them or to be able to easily turn them off or tune them. The recommendation engines are there to push Spotify's interests even when those interests don't align with the user. Bad UX and bad recommendation controls are in some ways a feature for the company, not a bug.
If the recommendation engine was presented in a way that consumers could easily control and if the majority of even casual users felt confident interacting with it and regularly turned it off, then that would be treated as a software defect to be fixed. Modern recommendation UX has more in common with advertisements or the popups websites put in front of content to force you to make an account, in the sense that you're not really supposed to take advantage of any opt-outs that the companies begrudgingly supply.
The problems you describe are why anything "algorithmic" is hated in some circles, despite there (IMO) being value in it in theory. People just can't imagine it ever being good because in reality it almost never is.
Also you can right-click a playlist and select "Exclude this from musical profiling" (translated from my native language). I did that for many older playlists I'm no longer into, and this seemed to cause a significant steer in weekly recommendations which have since turned to be much more in line with my current tastes.
Spotify should still be able to realise when something completely outside of the normal is played. Maybe as a different listening mode, eg "Kids party" - and keep them separate.
I had no idea this existed, nor would I know how to turn it on from Google Home / Alexa, which is my primary consumption method (particularly in "party" circumstances).
> I stopped paying for Spotify, and switched to Youtube, then Youtube Music. Youtube Music still does the good old thing, plays random songs staying within the genre I started with.
Perhaps only until they get the market position they desire...
Another good example of this is Apple’s “radio station” concept. It plays deep cuts unrelated to the originally chosen music about 3 songs into your list to spread the wealth (play songs with the smallest possible royalty costs).
Apple pays the same rate to everyone and for all music, and for what it's worth it is well known that Apple pays much more than Spotify does to artists.
It is incredibly difficult to guess what related music people will like, and some people will like the selections, others won't. Apple recently added a "Discovery" station that tries to guess music you might like but don't listen to and I am a big fan of it, while many others absolutely hate it. Eh.
Things like this are super annoying to me, especially when Apple is engaging in it as well. Just like deduplication on cloud services, this "feature" is literally only beneficial for the company so they can reduce your file storage use and make your data less secure and have an excuse to get to know more about your data they can use for other purposes.
> this "feature" is literally only beneficial for the company
Is this actually true? I don't know. I don't think consumer file hosting services would be viable at their present price points if they were storing every single file byte-for-byte.
Certainly some of the value capture belongs to the company, but if you're not particularly concerned about security it seems like a win-win.
If you are concerned about security, encrypt your files and everyone else with deduplicated files will subsidize the cost of your service.
It's not that they were simply smart and didn't de-duplicate identical songs; at some point they simply replaced instances of various recordings of songs with a single one. I could not be further from being an audiophile, yet I still understand the anger this could cause. Even at a basic level, there can be relatively major differences.
I don't think this has anything to do with de-duplication.
It's likely referring to iTunes Match which is a paid service that allowed you to upload your own MP3 files to Apple who would then serve it to you alongside any purchased/streamed music.
I believe that by default (but able to be disabled) your songs would be "upgraded" to the master-quality equivalent in Apple Music if they were identical. But then over time artists have been replacing those song with different versions/mixes.
> Just like deduplication on cloud services, this "feature" is literally only beneficial for the company so they can reduce your file storage use and make your data less secure
Every distributed object store I've ever seen takes your data, chunks it into a fixed size, hashes it and then distributes copies on multiple nodes.
So deduplication is intrinsic to the architecture and not some business decision layered on top.
I imagine they mostly all do it, with the exception of maybe Tresorit, possibly ProtonDrive but I am not super familiar with them. Its actually a "feature" of convergent encryption which Apple uses in lieu of true zero-knowledge e2ee like Tresorit.
> if ML is so powerful why does Spotify exclusively serve up the steamiest of dog sh$t?"
Has popular sentiment turned against Spotify’s algorithms? I thought their recommendation system was their most-respected feature – in my own experience at least, Discover Weekly continues to be Spotify’s moat.
I have pretty diverse music taste, and Spotify's discover weekly is nearly universally bad for me. My wife's taste is less broad, and hers is much better.
I listen to a wide range of music and Spotify's recommendations had been really good for me. But more recently they have gotten terrible. It's really annoying and has made the service much less useful.
I recently switched to Tidal, and decent recommendations are what I miss the most. While I prefer getting music recommendations from an actual person with similar taste, Spotify was generally good enough.
That's more than can be said for Tidal. My recommendation playlists generally contain tracks and artists I already listen to (sometimes from completely unrelated genres), sprinkled with what seems to be random tracks from the "more like this" section of said artists.
Still not enough to compensate for the shit stream quality and ever increasing price of Spotify though...
My range is pretty broad as well, and over all I'm extremely happy with Discover Weekly. It's probably my primary vehicle for finding new music today, honestly. But it does seem to be pretty hit or miss for other people, for some reason.
Lol what. Spotify has always been subpar on recommendations. Their moat is their brand, catalog and users. Discover Weekly is not even remotely a moat, it's trivial for anyone to make something similar or better. For me, discover weekly gives me less songs that I like than just turning on a good old radio channel
It's far too hidden to be their main feature. For a completely new user, it's not promoted as such. It's at best one of the many colourful "tiles" you can click on.
I've never noticed Spotify's recommendation engine is very good, but I came from Pandora.
What drives me absolutely bonkers is the number of foreign language songs in Spotify's recommendation engine for me. No matter how many times I say "never play this song again" to every foreign language song, give it about 5-10 more songs and it'll play another one.
I've never knowingly listened to a foreign language song
Cool story, but your friend knowing the answer has nothing to do with his ML education, I suspect, but with his business savvy. I say this because I used to work for a company in the online ads space; they were all knuckleheads but all of them would've told you the same thing.
People always like to figure a conspiracy. It is very unlikely Spotify is optimizing their discovery for profit. Their algorithm is one of their primary selling tools and while it isn't unheard of for brands to fuck their goodwill to increase profit, I just don't see it happening here.
Inevitably what ends up happening in these situations is that some product manager thinks, "wait, what if we had a 'gold' version that people payed for and that version also pushed whatever content we wanted at them?"
I was ranting, immaturely, to my ML educated friend recently about how terrible modern discovery algorithms are ("if ML is so powerful why does Spotify exclusively serve up the steamiest of dog sh$t?") when he explained this to me.
His response was "they aren't optimizing their discovery algorithms for your enjoyment, they optimize them for profit."