collaborative filtering + content based hybrid. offline: train embeddings (mat fact / 2-tower) nightly on spark. online: user embedding lookup -> ANN over song embeddings -> rank + diversify -> serve. classic two-stage architecture. they wanted me to mention exploration (epsilon greedy / thompson) explicitly

Loading...

Something went wrong while trying to load the full version of this site. Try hard-refreshing this page to fix the error.

Spotify NYC - recsys feb

collaborative filtering + content based hybrid.