Netflix Recommendation Engine: Which Signals to Use and How to Rank Them

Imagine you are designing the recommendation engine that decides which shows and movies Netflix puts in front of each member when they open the app. As the product manager, which parameters would you have the engine take into account when it chooses and orders titles for a member, and which of them would you weight most heavily? Ground the parameters in what the recommendations are meant to achieve for members and for Netflix, rather than listing every piece of data the service could collect. ```hint Sort the signals first Group candidate signals by what they tell you: something about the member, something about the title, or something about the moment the member opens the app. Then ask which group still works for someone who signed up yesterday. ``` ### Clarifying Questions - Which surface is in scope: the rows on the home screen, the title suggested when an episode or film ends, search results, or all of them? - What should the engine optimize for: hours watched, members finding something quickly, or members staying subscribed month after month? - Should recommendations be computed per profile, given that several people often share one account? - Are new members with no viewing history, and countries whose catalogs differ because of licensing, part of this design? ### What a Strong Answer Covers - A goal for the engine tied to Netflix's subscription model, with the member problem it solves (choosing what to watch) made explicit. - Signals organized into clear groups such as viewing behavior, explicit feedback, title attributes, session context, and household and region, rather than one flat list. - A reasoned ranking of which signals carry the most weight, including the difference between strong and weak evidence of interest (a finished season versus a tile scrolled past). - Cold start for members and titles, diversity, freshness, and catalog constraints such as titles leaving the service. - Success metrics with guardrails, and risks such as filter bubbles, popularity bias and privacy. ### Follow-up Questions - One profile is used by two people with very different tastes. How would the engine notice, and what would it do differently? - How would you weigh recommending Netflix's own productions against licensed titles a member is more likely to enjoy? - A new model scores better offline, but an A/B test shows no change in retention. What do you conclude, and what do you do next?