Collaborative Filtering: The Classic Engine

Collaborative filtering bets that preferences are redundant: your taste is predictable from people who overlapped with you. The whole apparatus lives on the user-item matrix — users down the side, items across the top, interactions in the cells, mostly empty.

Two neighborhood flavors:

• User-user — 'people like you liked…': finds your nearest neighbors in taste-space, recommends their favorites
• Item-item — 'items like what you liked': computes item neighborhoods once, looks them up at request time

Item-item won at Amazon scale for a structural reason: item neighborhoods are stable. Your taste drifts every session; 'people who bought this camera also bought this tripod' barely changes month to month — so it can be precomputed, cached, and served in microseconds.