Two-tower embeddings
User tower + item tower. Trained on watch history. At serve time, retrieve top-K items by embedding similarity. Sub-linear.
Advertisement
Two-tower embeddings
User tower + item tower. Trained on watch history. At serve time, retrieve top-K items by embedding similarity. Sub-linear.
Advertisement
Candidate gen narrows
15K titles → 500 candidates via embeddings + business rules (region availability). Fast.
Ranker orders
500 candidates scored by heavy DNN with 100+ features per candidate. Sub-100ms. Top 100 kept.
Row assembly is art
Homepage rows themed: 'Trending', 'Because you watched X', 'Popular in India'. Each row's items ranked separately.
Personalized artwork
Same title, different thumbnails per user. Preferences learned. Movie posters A/B tested by user cohort.