Item-Based CF
Predicts a rating from the user's own ratings of similar items, where two items are similar if the same users tend to rate them similarly (measured on the columns of the user-item matrix, not from item features).
When to use
Lecture 6 shows a worked Cosine Similarity between two items and why item-based often beats User-Based CF: items are stable, the item-item matrix is smaller and can be precomputed. Amazon's "also bought" is item-based CF.
See Lec 6.