Spotify: Building Music Recommendations at Scale
How Spotify built a recommendation engine that analyses billions of listening sessions to predict what 600 million users want to hear next.
The challenge
By 2015, Spotify had hundreds of millions of songs and users who wanted personalised discovery, not just search. Generic 'most popular' playlists felt impersonal, and manual curation couldn't scale to millions of unique listeners. The technical challenge was building a system that could analyse listening patterns across an enormous catalogue and surface genuinely relevant recommendations in real time, not batch-processed overnight.
The strategy
Spotify combined two distinct machine learning approaches: collaborative filtering (finding users with similar taste and recommending what they liked) and natural language processing on text written about music — blog posts, articles, forum discussions — to understand genre and mood associations that raw listening data alone couldn't capture. This hybrid approach became the foundation for Discover Weekly.
Want the full story, outcome and quiz?
Read how Spotify executed it, the results, key lessons and test yourself with a quiz. Free on CaseLearn.
Try the full case free →Key lessons (preview)
- Combining structured data (listening history) with unstructured data (text about music) captures signals neither source reveals alone.
- Deliberately constraining a feature (weekly, 30 songs) can increase perceived value more than making it real-time or unlimited.
- Recommendation systems are a product design problem as much as a machine learning problem.