How Collaborative Filtering Predicts What You'll Like Without Knowing Content

Collaborative filtering is a recommendation technique that predicts user preferences using only patterns in ratings data, requiring no knowledge of item content such as plot or genre tags. The method begins with a sparse user-item ratings matrix, where the goal is to accurately fill in the missing values and surface top predictions for each user. Two classical approaches exist: user-user filtering, which identifies users with similar rating histories, and item-item filtering, popularized by Amazon, which groups items rated similarly by the same people. Item-item filtering is generally preferred in production because item relationships are stable over time and can be precomputed, making recommendations faster and more reliable. Matrix factorization, the technique behind the Netflix Prize-winning solution, takes a broader approach by learning hidden factor vectors for users and items whose dot products approximate observed ratings.