Recommender Systems

A recommender fills in a huge, mostly empty table: users down the side, items across the top, and a few known ratings or clicks. The job is to predict the blanks.

Collaborative filtering uses only the table:

• Neighborhood methods: find users with similar tastes (or items liked by the same people) and borrow their preferences, a form of nearest neighbors.
• Matrix factorization: give every user and every item a short vector, an embedding, such that the dot product of a user’s and an item’s vectors predicts the rating. The dimensions often end up meaning something, like “comedy versus drama”, without anyone defining them. This approach was central to the Netflix Prize competition (2006 to 2009).

Content-based methods use item features (genre, description, audio) and help with brand-new items. Modern platforms combine both in large neural models with many signals, and must balance engagement with variety and wellbeing.