Streaming and Video Recommendations

In October 2006, Netflix offered $1 million to anyone who could beat its movie-rating predictions by 10%. The Netflix Prize drew thousands of teams and was won in 2009 by a merged team, BellKor’s Pragmatic Chaos. Its most lasting lesson was the power of matrix factorization, a form of collaborative filtering:

• Arrange ratings in a giant user × movie table, mostly empty.
• Learn a short vector (embedding) for each user and each movie so that their dot product predicts the rating.
• Similar tastes and similar movies end up close together, with no one telling the model about genres.

The winning solution blended hundreds of models (ensembles), and Netflix never deployed it in full: the engineering cost was too high and the business had shifted to streaming.

Today’s recommenders at Netflix, Spotify and YouTube typically work in two stages:

1. Candidate generation: quickly find a few hundred plausible items out of millions, using embeddings and vector search.
2. Ranking: a neural network scores each candidate using many signals: history, time of day, device, freshness.

Every change is tested with A/B tests. The hard questions are about objectives: maximizing clicks or watch time can create filter bubbles and reward sensational content, which is why diversity and user satisfaction are measured too.