Live data from Hacker News

Spotify’s Discover Weekly: How machine learning finds new music

hackernoon.com

231–240 of 275 posts

Re: Spotify’s Discover Weekly: How machine learning finds new music

#231
My Foolproof Guide To Managing Your Discover Weekly Playlist:

- Only listen to the DW playlist once through. Find the songs you even remotely like, put them in a new playlist. Listen to that playlist instead (I call these playlists "DW-{datestamp}").

- Find your favorite songs in the playlist, explore that artist/album. Even if you don't think you'd like the other stuff listening to more of an artist seems to help suggestion variety.

- Don't let your listening for the week be dominated by a good Discover Weekly playlist... every time I do this my next 2-3 weeks are total crap. If you must repeat the same songs over and over, move them to a new playlist.

- Try to mix up genres as much as possible... listening to different genres that aren't your favorite often leads to Spotify recommending off the wall artists in the genres that are your favorite.

- Keep a playlist of your most frequently listened songs, regardless of genre, artist, etc. Whenever you want to listen to one of favorites listen from that playlist, instead of going to their artist page. For some reason this seems to have a large effect on my recommendations.

Most of these suggestions are due to personal experience and two theories:

1) Grouping/organization/total play count of playlists influences recommendations much more than people think

2) It's very easy for Spotify to get into positive feedback loops, forcing variety and constantly curating/making new playlists expands your horizons and keeps the recommendation engine from repeating songs/artists too much.

There seems to be a common misconception, even among programmers who should know better, that Spotify will just instantly and always offer you fresh, undiscovered, music you like on demand... which isn't how these systems work. It definitely makes it easier to find new music, and is a valuable tool in finding new artists or under marketed artists, but you're still going to need to put in a modicum of work curating your own experience to get the most out of the discover playlists.

Re: Spotify’s Discover Weekly: How machine learning finds new music

#233
post #226

Earlier quoted context omitted.

The importance of a good Discover Weekly has caused me to be super 'careful' with what I listen to. I have a separate Spotify account on my family plan where I listen to stuff I don't like. For instance, it's connected to my Sonos so guests at my house can listen to whatever they want.

Something like an "Incognito Mode" would be a good idea for Spotify and similar services.

They actually have one. "Private listening". But you can't assign certain devices to stay in that mode.

Re: Spotify’s Discover Weekly: How machine learning finds new music

#234

Earlier quoted context omitted.

>1. I hear repeat songs in my discover all the time Something is wrong. Discover Weekly is supposed to have only songs you have never listened to, even once. That's how it works for me.

I get repeat songs now and then. Not a lot, but a noticeable number.

Perhaps you've listened to the regular version of a song, but the Discover Playlist has surfaced its "radio edit" or some other very, very similar version?

Re: Spotify’s Discover Weekly: How machine learning finds new music

#235
post #226

Earlier quoted context omitted.

The importance of a good Discover Weekly has caused me to be super 'careful' with what I listen to. I have a separate Spotify account on my family plan where I listen to stuff I don't like. For instance, it's connected to my Sonos so guests at my house can listen to whatever they want.

Something like an "Incognito Mode" would be a good idea for Spotify and similar services.

and exists: WIN: File/Private Session Android: Your Library/Settings/Social/PrivateSession

This will cause your songs not to be influencing the AI curated playlists (like Discover Weekly and Daily Playlists)

Re: Spotify’s Discover Weekly: How machine learning finds new music

#236

As an artist, Discover Weekly has been the best thing to ever happen to me. Every Monday I get a big infusion of listeners (around 5,000)— many of who stick around and check out my other music :) Prior to that, the best press I could get was the tedious process of cold-emailing bloggers (a practice which is now dying off).

Never thought I would find a new artist on hacker news that I would like. Purification ritual is great!!

That's amazing— thanks!

Re: Spotify’s Discover Weekly: How machine learning finds new music

#237

Earlier quoted context omitted.

Wow, what a great user experience.

If you had it all on screen you would talk of button fatigue. Users are the worst.

Did you know Spotify also has a its approve and disapprove buttons in the desktop app on opposite sides depending on what particular radio-like feature you're using?

Re: Spotify’s Discover Weekly: How machine learning finds new music

#238

I find that my music tastes can vary too much for Spotify. While working, I mostly listen to Electronic music. But I can also take a hard right turn and listen to Hall and Oates... or Fleetwood Mac, or Steely Dan. Then some form of 80's throwback. But as far as recommendations go, I really only want to expand in that first genre of Electronic Music. I don't want Jethro Tull to be recommended to me. My feed used to be…

This is the problem I have had with it. Last month I decided to check out some Latino music from Columbia after a recommendation, and my Discover Weekly playlist after that started to get populated with almost all Portugeze songs and Brazillian artists. I did find some great songs in them too, but I would have preferred more Latino music. It was disappointing that their model is not able to properly disambiguate between two popular, yet distinct languages/genres.

Re: Spotify’s Discover Weekly: How machine learning finds new music

#239
I built the foundation of this system while at Spotify. While it's true that we looked at a lot of different signal, at the point when I left (early 2015), it was all based on collaborative filtering.

The reason collaborative filtering works so much better than anything else is that given enough data, it will already encompass everything else. If there are reasons why certain users prefer certain sounds, or certain lyrics, those patterns will emerge in the listening data.

The main reason to use any non-CF method is mainly for new content that Spotify doesn't have much listening data for.

I'm no longer at Spotify, but let me know if you have any questions

Re: Spotify’s Discover Weekly: How machine learning finds new music

#240

It's very impressive. But, I'm curious about some of the technical details. For each of these representations you basically end up with a dense vector on a song level. Which I assume you would then kNN with a user specific vector. But I've never come across a nice kNN data structure that supports high dimensional vectors in a larger than memory setting whilst supporting updates. Spotifys own Annoy is cool https://git…

Spotify re-runs the latent vector models regularly and re-indexes them into Annoy indexes. There is no need to do that in real time, you can be a few weeks delayed and it's usually fine. New music doesn't have much data and need different methods anyway.
Post reply on HN