Video recommendation

How this question breaks down

1. 1Candidate retrievalThe funnel6 minNarrow hundreds of millions of videos to a few hundred worth ranking, using two towers and a nearest-neighbour index.
2. 2RankingThe funnel24 minScore ~500 candidates with one multi-head model that predicts click, watch, satisfaction and dismissal, correct the labels for video length and screen position, and blend the heads into expected valued minutes.
3. 3Re-rankingThe funnel20 minTurn the ranker's sorted list into a page: filter what must never show, spread out near-duplicates with MMR or a determinantal point process, keep the viewer's mix of interests, and let new uploads from followed creators in.
4. 4Cold startEdge cases20 minA new video has a title, pictures, sound and a creator but no watches; a new viewer has a phone, a place and a first few taps. Build from what exists at minute zero, learn fast from the first signals, and measure the new slices on their own.

Work through the parts in order and you’ve covered the full design. Each part is one topic.