User-Based CF

Predicts a user's rating for an item from what similar users gave it: find the most similar users by cosine similarity over the items both have rated, then take a similarity-weighted average of their ratings.

When to use

Used in worked rating-prediction questions (User A to I8 came to about 2.8). It is the same idea as KNN, with K neighbours voting on a rating instead of a label. Weaknesses: thin overlap gives unreliable similarity, it is slow with millions of users, and a user with zero ratings cannot use it. Mean-centering corrects for users who rate on different scales.

See DS Lec 6.

Appears in

No references yet.