"People who rated like you also rated…" Find users similar to the target user (Alice), then borrow their ratings for items Alice hasn't seen. The top-k similar users vote on missing ratings.
User-item rating matrix (target = Alice). Top-k most similar users highlighted in green.
Alice's row (target)
Top-k neighbors
Predicted rating for Alice
The mechanic: compute cosine similarity between Alice's rating vector and every other user's. Pick the top-k. For each item Alice hasn't rated, predict her score as the similarity-weighted average of those k users' ratings on that item. Higher k = more stable but blurrier.