Show HN: Selecta – Tune Your Own Spotify Recommendation Algorithm | Hacker News Reader

Show HN: Selecta – Tune Your Own Spotify Recommendation Algorithm

selecta.ciaran.co.za

69 pointsparabyl31 comments

I posted here a while back with this app I'd been working on. In short:

It allows you to talk to Spotify's recommendations API to specify musical features you'd like recommendations similar to.

It's loads of fun to mess around with, and I've given it an overhaul to include many more dimensions ito. musical features, as well as the ability to log in with Spotify, and save things you find to a playlist. You don't need a subscription to Spotify to use the app either.

Thanks to @Mockapapella for the tagline when I posted the first version: https://news.ycombinator.com/item?id=35907215

Open articleView on HN

Interesting! Too late for me; Spotify's declining recommendation capability was one of several issues that allowed me to cancel my subscription.

I really wish streaming services would just let us give conscious direction instead of trying to hack our brains. They're not as good at as they think.

If you still want to give it a shot, you shouldn't need an active subscription for it to work. But I do totally get moving away from the streaming world. Even bandcamp is losing its ground as a safe a haven for making sure artists are paid their worth.

Wait, what changed at bandcamp?

it got sold to Epic Games a while back.

Ah thanks

yeah this is what I was referring to, thanks

Did you end up switching to a different platform? If so, which one?

I continue to use YouTube, which is less dynamic (meaning, you can end up on an actually or nearly deterministic playlist starting at a given song, rather than on "random" songs), but really surfaces some incredible stuff. It does take more work to break out of a mold sometimes, but I tend to find more interesting things when I do. Plus every song is on there.

I also began using Tidal and I'm unsure if that's going to be a long-term thing for me. It's good so far but I find I'm not using it as much; I think it will take effort to get my library into its algorithm before it becomes really useful.

One problem with Spotify is it has a very poor memory. It would only focus on things I had recently listened to and almost never accounted for my full library of "liked" songs/albums when giving suggestions (so something like Pink Floyd, despite having many liked songs, would absolutely never appear on a generated playlist). I remember Pandora was even worse at this; I don't know why. Anyway it's too soon yet to tell if Tidal will do the same, but I don't have high hopes. Even YT tends to do this. It's extremely frustrating.

I find youtubes music algorithm to be on point most of the time for me. I just click one song and listen bunch of songs i enjoy, new and old.

Completely opposite experience. It quickly drifts either its own 'genre' playlist or your own history tops or nowhere. I find it much easier to pick some community playlist and it has more chances to be self consistent than any 10 radio songs played by yt music.

Seems to ask for quite a lot of data:

> Your email > The type of Spotify subscription you have, your account country and your settings for explicit content filtering > Your name and username, your profile picture, how many followers you have on Spotify and your public playlists

> Take actions in Spotify on your behalf > Create, edit, and follow private playlists > Create, edit, and follow playlists

Hey, thanks for the feedback. Sign-in and saving to a profile was the most requested feature across the board when I previously launched Selecta, and as such it needs access to enable features that were requested by those who enjoyed using it.

I'd be more than happy to share what each of those datapoints is used for (and, in some cases where I only need an auth scope for one permission, but they're grouped together). Selecta is strictly open-source and anonymous analytics.

You can find the auth scopes I use in this file: https://github.com/knightspore/selecta/blob/main/src/app/api...

and the related access they provide here: https://developer.spotify.com/documentation/web-api/concepts...

A feature I've long desired would be to be able to ask what lies between or along the average or sum of the vectors representing two arbitrary songs or bands. For example what music is midway between "Florence and the Machine" and "Primus"? Or, what happens if you take the vector from Green Day to Modest Mouse and extend it - do you land on another (presumably obscure indie) band?

If Spotify won't give me those toys, are such things possible via the API that Spotify has provided?

There's a webapp called Boil the Frog that does exactly this!

Cheers, going to give it a look.

This is very cool, hadn't heard of it before. Thanks.

I mean... Spotify tracks a bunch of cool (algorithmically measured) sets of metadata for each song. Basically the ones you have on this site. And there are APIs to retrieve them for any song, and you can use them all as arguments for search. I don't see why you wouldn't be able to find a thing in between two things.

Keep in mind that the actual accuracy of those metrics is debatable. They are cool... But mostly for funzies, or for approximate sorting in 3rd party tools.

If I understand correctly, what I'm getting back from their API is basically what lies between, but without any of the interesting data they're using the do the calculation. One can hope, but I'd imagine that data is priceless to them and won't ever be available to us.

I have been playing with this and I really like the concept. A few nitpicky points of feedback

1) It would be nice if the button to play a sample of the song was located on each song or somewhere close to it, so that you wouldn't have to move your mouse to the bottom

2) The volume slider isn't super obvious as to its function, I originally thought it was a seek bar. Maybe a Volume icon would help here

3) Most importantly, it would be great if the app could play full tracks for users with Spotify Premium, preferably automatically as well, so that this could be useful as a "radio". One feature I really want is to have a "tunable recommendation radio" of songs that all sound a certain way and this comes very close.

Glad you're enjoying it, always open to nitpicky feedback. I'll put all of these suggestions on the backlog and see how they play together - thanks!

I will try this out later, I find spotify recomendations a pure torture. The best I got was old last.fm recomendations, while they still had streaming, then soundcloud and pandora. The best ones had dislikes, not just likes.

They really can be terrible sometimes. I loved last.fm too - still sporadically scrobbling at https://www.last.fm/user/Knightspore. But today my profile mostly serves as a record of music I had come across more 'organically' prior to mainly consuming via streaming services.

It would be great if the tool could load your liked songs

Thanks for giving it a try - I'll take a shot at that one.

I'm not sure if I've just already found all the good music released in the past few years by my 10 favorite artists or what, but the recommendations I got were either songs I didn't like or songs that Spotify already tends to recommend me (which are great, but I already have them all saved)

This is great! I've done a similar thing with knn clustering and it had fun results. Would you mind sharing a bit about the data science behind this project? One issue I ran into was the numeric features I clustered with didn't care about genre, so the resulting playlists had some interesting mixes.

Thank you! That's super cool - I'm not nearly as familiar with data science as it sounds you are, I've just made some choice placements in building a UI around the Spotify API. They're hiding the nuts and bolts of turning a set of music feature parameters into a list of recommendations.

I got a message that says "Try signing in with a different account".

maybe give it another shot? I wasn't able to replicate the issue but haven't seen anyone else have it

Great discovery tool!

Thanks! Glad you're enjoying it :)

Reply on news.ycombinator.com