A recommender or personalization model has nothing to go on for new users or new items that have no interaction history yet. The usual fixes use content features, popularity and onboarding questions until enough interactions arrive.
Card 1 of 228. Answer: A recommender or personalization model has nothing to go on for new users or new items that have no interaction history yet. The usual fixes use content features, popularity and onboarding questions until enough interactions arrive.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
Collaborative filtering learns from interactions: who clicked, watched or bought what. A new item has no interactions, so it has no useful embedding; a new user has no history to match. A brand-new platform has both problems at once, sometimes called system cold start.
Example: a streaming service uses matrix factorization. A film added today has no ratings, so its latent vector is meaningless and it never ranks high enough to be shown. Because it is never shown, it never collects ratings, and the problem feeds itself. The fix is to start its embedding from metadata (genre, cast, a text embedding of the synopsis) and reserve some traffic for exploring new items, for example with a bandit.
For new users: ask for a few preferences at sign-up, use context such as device, country and referral source, and fall back to what is popular in their segment. Two-tower models help because the item tower can embed a new item from its features alone. Interviewers also ask how you would measure it: report metrics separately for new users and new items, since averages hide them.
