So far, we have compared users to other users in order to make our predictions. However, the similarity space can be partitioned in two ways. User-centric collaborative filtering plots users in the preference space and discovers how similar users are to each other. These similarities are then used to predict rankings, aligning the user with similar critics. Item-centric collaborative filtering does just the opposite; it plots the items together in the preference space and makes recommendations according to how similar a group of items are to another group.

Item-based collaborative filtering is a common optimization as the similarity of items changes slowly. Once enough data has been gathered, reviewers adding reviews does not necessarily change the fact that Toy Story is more similar to Babe than The Terminator, and users who prefer Toy Story might prefer the former to the latter. Therefore, you can simply compute item similarities once in a single offline...