A content recommendation engine is the software inside a streaming service that decides which films, shows and matches each viewer sees first when the app opens. It chooses the order of the rows on the home screen, and the order of the titles within each row. It bases these choices on what the viewer has watched before, and on what people with similar tastes watched.
Why is it needed? A large service may hold tens of thousands of titles. A TV home screen shows perhaps fifteen at a time. Most viewers decide what to watch within a minute or two. If nothing on the first screen interests them, many close the app. So the choice of those fifteen titles matters a great deal.
The engine works in two stages, like most recommendation systems. In the first stage, it quickly picks a few hundred candidates out of the whole catalogue. It uses simple rules, such as "titles watched by people who watched what you watched" and "new titles in your language". In the second stage, a machine learning model scores each candidate. The score estimates how likely this viewer is to start the title and keep watching it. The highest-scoring titles go first.
Say a retired engineer in Nagpur has watched two Marathi thrillers and a cricket documentary this month. He opens the app on Friday evening. From a catalogue of 40,000 titles, the engine picks 500 candidates. The model scores all 500 in a fraction of a second. His top row now shows three more Marathi thrillers, a new Hindi crime series and a documentary on a famous Test series. The engine even picks the poster for each title. For a film with both a car chase and a family scene, he is shown the car chase poster, because he chose action scenes before.
The engine also follows the service's own rules. During a live match, the match is placed at the top for everyone. A new film the service paid a lot for may get a guaranteed place in the second row. Titles that are not allowed for a child's profile are removed.
The engine needs data about each viewer. The streaming apps record every title started and every title abandoned. Some services also join this with other customer data in a customer data platform (CDP). In India, viewers often share one account across a family. So the engine must guess which person is watching from the time of day and the titles chosen, unless each person uses a separate profile. This part belongs to a streaming service, and the same idea ranks short videos and posts in a social media app.