How Spotify recommends a song nobody has played yet

transcript

This song came out this morning. Nobody has played it. So how does Spotify know to recommend it to you?

The classic trick: users like you also liked this. A taste vector per user. And a taste vector for every song. 128 numbers each. Learned from what people play.

But here's the catch. It learns from plays. A brand-new song has none. To this model, it doesn't exist.

So Spotify listens to the song itself. Its spectrogram. A neural network reads it. Tempo, energy, mood. Out comes its own 128-number vector. A song with zero plays now sits beside sound-alikes.

Collaborative filtering, for users like you. Content signals: tags from blogs, reviews and titles. Audio analysis, for songs that sound similar. Weights come from A B tests.

Then Discover Weekly. 30 tracks. 24 are top scores. Safe bets. 6 are deliberate exploration. Outside your comfort zone. That's how new music collects its first plays. And no filter bubble.

No plays? Listen to it. Then let people find it. Three models beat one.

Full breakdown in the lecture.