Discovery level in personalized playlists

[eduluiz]

I would love to have a Discovery level option on Spotify.

Currently most of the songs on Daily Mixes or Autoplay are songs that I already liked. This is great but sometimes I am under the impression that the same songs are always shown to me and I wish to have the option to adjust the Discovery level, for example:

- High: high discovery would show more unknown songs that I could like, let's say around 70 percent of unknown songs and 30 percent liked ones.

- Normal: around 50% of liked/unknown songs.

- Low: this seem to be the current way Spotify works, showing around 70 or 80 percent of liked songs.

The new "Discovery level" option would influence Daily Mixes, Autoplay and Radio stations once defined by the user. The default option would be how things work today, but it would be very helpful give the users an option to define if they would like to discover more music.

I hope you may consider this idea and hopefully implement it soon.

[Daserr]

The spotify-algorithm is great but it is tailored to a broad audience to work for everyone. If you are looking for specific music recommendation-patterns or try to circumnavigate situations, in which the algorithm is stuck in a direction you dont want to go in, you are out of luck under the current system.

I suggest implementing a UI that allows users to adjust the statistical model of the algorithm. I have two primary options in mind:

1. Focus Width from narrow to broad
2. Number of neighbours from high to low

By focus width I mean the statistical deviance the algorithm is set to. A lower value would lead to results that stay closer to the source material (Genre, artist, other songs listeners like) and a higher value would lead to more variety but less correlation with the source-data.

By number of neighbours I mean the number of songs in for example a playlist that are accounted for to recommend a song. A lower value would mean that a smaller set of songs leads to any individual recommended song, which produces songs that closely match each just a few source-songs. A higher value would take a larger set of songs into account for any single recommendation.

There are two more related options I would like to propose:

1. Change direction
2. Forget timeframe

Change direction is similar to reload under playlist recommendations but doesnt just load a new batch, but also changes the source-parameters taken from the playlist a bit, to yield different results

Forget timeframe is quite self-explanatory and is just an option to not take listening-behavior from a set timeframe into account anymore. This allows users to listen to music they dont want recommendations for without having to remember to activate a private session. It also allows user to act retroactively.

At last I want to recommend an option to import a friends algorithm settings, if they share it with you, so you can change to their algorithm in the settings and experience the recommendations as they do.

[hans-jürgen]

It seems that e.g. the number of followed artists influences the number of listed songs on your Release Radar playlist, so adding more artists to your Spotfy profile might already help:

https://community.spotify.com/t5/Desktop-Windows/199-new-songs-on-Release-Radar/m-p/4906876#M84081

[Daserr]

Thank you for commenting! I am already doing that and it does work. But it doesnt really allow me to do what I have suggested here. Still a good tip! It's useful to use a tool that auto follows all artist you have songs saved of.

[JanBrecht]

I suggest implementing the idea gnoosic.com uses. This website allows you to select three of the bands you favor at that moment and bases recommendations on them. This way a more personal algorithm can make a playlist.

[Daserr]

Gnoosic.com is brilliant, but I actually sent them a feedback message with pretty similar things to what I have said here. 😄

[Katerina]

Updated on 2020-03-06

Hey there @Daserr,

Thanks for sharing this! Before we give the proper status to your idea, we'll just need a bit more info to understand it first.

You mention that you'd like a UI that allows users to adjust the statistical model of the algorithm. Could you give us some more info on how that would look/work in the app? For example, if the options you mention show up when playing certain playlists, if they would be adjustable to different values each etc?

We'll be keeping an eye out for your reply, thanks! 

[Daserr]

Hello @Katerina,

thanks a lot for noticing my idea! I'd say there are three areas where an implementation of the algorithm adjustment idea would make sense at first: The recommendations at the end of playlists, discover weekly and release radar. This way, users may explore different configurations without influencing their overall-recommendations.

Akin to the "reload" button above the playlist-song-recommendations (I dont exactly know the name in the english UI, but I believe it's also called "Continue listening") there could be one more button called "adjust". A click on that button would expand a menu with two sliders for focus width and number of neighbours, and the "change direction" button  with a small one-sentence-explanation to what these adjustments mean. Of course there also needs to be a button to revert to standard values.

For discover weekly and release radar, the adjust button could be placed next to the "Play", "Heart" and "Three Dots" elements.

I'd also propose a new settings-tab called "Discover" that offers the "Forget Timeframe" option and the option to share and load algorithm-configurations (There could be an expanding menu, in which you could click on either your own settings, or the ones the people you follow shared publicly or with you). I wouldn't recommend to implement the other options here, as people could "mess up" their overall algorithms, its better to keep this confined to individual playlists. 

I'd love to see this idea or something similar implemented, as one of the most important things about spotify is finding new music (At least for me). Also, recommendation-algorithms across all platforms are currently still stuck in a phase, in which nobody in the general public knew about machine learning, and the elegant way to implement them, was to give the user very limited feedback-options ("I like this"/ "I dont like this"). Now we know how they work (at least generally) and I dont want to have to "train" my algorithm myself just to change minor things. True personalization means, being able to adjust the "personality" of the system recommending me stuff. I want my algorithm to be more adventurous and less scared, my friend wants his algorithm to stay focused and dont drift to far away from his input-songs. These are drastically different approaches and with the current spotify-algorithm we have to meet somewhere in the middle.

The tastes in music are as diverse as people are, and spotify does a good job in giving everyone their own experience, but on the meta level we have no options. It's like having a friend, who's always recommending you music, but you only have that one friend to ask.

[Katerina]

Updated on 2020-03-06

Hey again and thanks for adding these details @Daserr, 

We've marked this as a new idea - we've also edited slightly the title of your idea to make your suggestion more clear to other users.

Spotify Staff will look into this idea once it reaches the necessary amount of votes. More information about how the Spotify Idea Exchange works can be found here.

[TheRealSlimSaad]

If I listen to radio based off one of my playlists, let me prevent songs from that playlist from showing up in radio.

Sometimes I'm just looking for new music, and if I have a playlist of ~100 songs the radio will assuredly have at least half of them.

[osornios]

Updated on 2020-04-23

Marked as new idea, which is a repost of this inactive idea.