SwedenMedia and entertainment2025 – ongoingSpotify
Shared user embeddings for recommendations and search
Spotify reported 13% more item-to-stream conversions in tests of shared user embeddings
Problem
Spotify needed recommendation and search models to respond to changing listener preferences. Separate feature engineering across products duplicated work.
Approach
Spotify built an autoencoder that compressed listener signals into shared user embeddings. Product teams adapted these embeddings through lightweight transfer learning. The production system combined batch processing with updates triggered by listener activity.
Results
Takeaway
Sharing embeddings also required coordinating model updates so that search and recommendation systems continued to interpret them consistently.
Categories
- Industry
- Media and entertainment
- Business function
- IT and R&D
- Innovation type
- Process innovation
- Adoption
- Built in-house
- Evidence
- Company's own claim
Sources
Source summaries
Spotify described its deployed embedding framework and coordinated model updates. It reported 13% more item-to-stream conversions and 2.9% more discoveries in Home candidate-generation experiments.
Full details are in the original sources.
