It's absolutely wild. I have over 50 playlists I've made, some over 10 hours long with no repetitions anywhere. I have extremely thorough and diverse music interests and habits d

And yet, when Spotify's shit tier algorithm takes over, it kicks me into a similar 5 to 6 meme set of songs every single time. It's an absolute joke.

Same. Out of 5,000 songs, am I really supposed to be hearing certain songs 3+ days in a row, when I'm only listening to 20 songs a day...?

I can't tell if it's:

- Certain artists are paying Spotify to favor them in randomization?

- There's some kind of shared random seed across devices that results in picking a tiny subset of songs to randomize from in the first place?

- Other?

I do notice that the effect seems to persist for maybe a week, then I'll never hear those songs again, but now it'll be different songs that keep popping up repeatedly.

There's a related effect when you launch a radio station based on an artist or track. If you launch it multiple times in the same day or week, you get the exact same list of tracks. But maybe a week later the tracks have changed, like the radio has been recalculated based on a different random seed.

There's maybe some counterintuitive math going on here.

If you have 100 songs and listen to 1 song per day (which is 1% of the library), on any given day your odds of hearing the same song as yesterday are 1 in a 100.

If you have 1000 songs and listen to 10 per day (still 1%), the odds of hearing a song that was also played yesterday are a little less than 1 in 10, right?

So what matters is not only what fraction of your library's play time you sample daily, but also how finely subdivided the time is into individual tracks for sampling.

A bit of birthday paradox too? :)

> If you have 1000 songs and listen to 10 per day (still 1%), the odds of hearing a song that was also played yesterday are a little less than 1 in 10, right?

No. It's 10*(1/1000)=1%.

It'll happen a few times a year only.

The odds of hearing a specific song that was played yesterday is ~1%, the odds of hearing any song that was played yesterday is ~10%.

What are the chances of hearing the same three or four every day for a week though out of a set of a thousand? Certainly that must be quite unlikely.

No it's not. That's already accounted for by the "10*".

It's 1%. Any specific song is 0.1%.

That's not quite right. You need one 10 for the fact that you listened to ten songs yesterday, and another 10 for the fact that you're listening to ten songs today.

Assume that you got 10 unique songs yesterday, which is the case ~96% of the time. Then there are 990 songs you didn't hear yesterday, and for every song you listen to today, there is a 990/1000 chance that it's one of those songs. Hence the chance of only hearing new songs today is (990/1000)^10 = 90.4%.

I was excited to get an insight into some hidden backend tools but it seems like this post is just about playlists (creating your own playlists, paying for placements in other people's playlists), and ads?

Out of curiosity, does it happen on some clients and not others?

I remember hearing of a bug where if you played on a remote device, it would transfer the first part of your playlist (10? 100 tracks?), and then shuffle would only choose from among them.

But it's been 5+ years so things may have changed and/or I could be remembering completely wrong.

That might be true, but my impression is that the algorithm is weighted, so some artists (more popular in general, recommended, played more often by that individual user, would get picked more frequently by the "random" algorithm).

I wonder if this is market manipulation and not a honest mistake. To make certain labels and artists more money.

As far as I can tell all Spotify's playlists (in all forms, e.g. including "random" and "track radio") use weighted algorithms based on several factors like user's play history, Spotify's recommendations (probably includes paid promotions), general artist's popularity, etc.

When I still used Spotify, I would get a dozen of my favorite artists mixed into basically any "playlist" I pick. Was one of the reasons I quit Spotify - they are too opinionated on what I should listen to.

It also seems to put anything I added recently on heavy rotation.

My "discovery" section is full of albums I have liked + downloaded. It's infuriating.

The effect is so strong that 3 of those songs become my top 5 listened despite me not actually wanting to listen to those songs THAT much

It might think you really like those songs. I listen to my favorites every day, among others.

Eh, unexpected clustering is sort of okay but again, it's what people respond to versus what they think they will. I've written some scripts so I can dredge playlists off of some radio stations that are hooked up to the Internet as part of a quixotic little project of mine and I've gone through, looked for dupes, etc. Let's say we're doing new wave (sure to start an argument there). What people seem to dislike is ABC, The Buggles, the Cure, Duran Duran, Ebn Ozn, Fra Lippo Lippi, Duran Duran, etc. Just having a gap between there feels insufficiently random.

Clustering apparently ought to feel deliberate. Now think back to when you had actual DJs selecting tracks on the radio. One of the techniques was "Two from a particular band." Not two from a band with some tracks between them.

Similarly, one can do a "Four tracks from 1994" to provide a cluster in time, another technique I've heard.

If anything, the more metadata you have, the more you can provide short runs of something. Microgenres, for example.