Problem
A recommender treats every stream as a statement of taste. That fails when people play music to sleep, work out or learn a dance, when they play it for kids or guests, and when several people share one account. The recommendations then fill up with sleep music or children's songs, and the real interests rarely show. At Spotify, an analysis found millions of users with this "functional" listening, and they used recommendations less than users without it.
Example from the article: the author's own history showed more than 30% of his listening was sleep music, so his Discover Weekly was full of sleepy piano while his daily upbeat EDM and funk almost never appeared.
Solution in brief
Let users mark a playlist as "excluded", and apply the flag inside the shared taste-profile systems rather than in each recommender. All plays from that playlist, past and future, are then dropped from the profile. Surfaces that need factual history can still read the unfiltered data.
Questions this case answers
- How to keep implicit-feedback history that does not reflect a user's preferences out of every model that consumes it, without changing each downstream model
- How to validate a partially built personalization control with a live experiment before the full backend work is funded
- How to stop background, shared or other-people listening from distorting a user's personalized recommendations?
- How to give users control over what feeds their taste profile without changing every recommendation model?
Tools and methods
- Playlist-level exclusion flag — a user-set rule that removes plays from a chosen playlist from the taste profile.
- Central taste-profile layer — recent listens, top tracks and artists, user embeddings and interactions, where the filter is applied once for all surfaces.
- Painted-door prototype and A/B test — a message on about 100 editorial playlists plus a partial backend to prove impact before full build.
Google or a one-time email code. No password.