Block 7. Recommendations and Personalization

On a big catalog, viewers watch what the system surfaces. How streaming personalization — recommendations, search, merchandising — decides retention.

How a video recommendation system works: candidate generation and ranking, collaborative vs content-based filtering, cold start, and watch-time metrics.

Content metadata is the fuel for discovery: the four metadata types, EIDR and MovieLabs MEC standards, Schema.org, and the pipeline that surfaces titles.

How streaming search works: keyword vs semantic vs hybrid search, typo tolerance, voice and TV search, and the zero-result handoff to recommendations.

How streaming home screens are personalized: row selection and ordering, per-viewer artwork via contextual bandits, and editorial-vs-algorithmic balance.

How streaming services A/B test: watch time and retention as the OEC, the peeking trap and sequential testing, and interleaving for rankings.

How streaming personalization pipelines work: event collection, real-time vs batch, the feature store that kills training-serving skew, and VPPA limits.