Item-item collaborative filtering

Earlier collaborative filtering systems based on rating similarity between users (known as user-user collaborative filtering) had several problems:

• systems performed poorly when they had many items but comparatively few ratings
• computing similarities between all pairs of users was expensive
• user profiles changed quickly and the entire system model had to be recomputed

Item-item models resolve these problems in systems that have more users than items. Item-item models use rating distributions per item, not per user. With more users than items, each item tends to have more ratings than each user, so an item's average rating usually doesn't change quickly. This leads to more stable rating distributions in the model, so the model doesn't have to be rebuilt as often. When users consume and then rate an item, that item's similar items are picked from the existing system model and added to the user's recommendations.

Method

First, the system executes a model-building stage by finding the similarity between all pairs of items. This similarity function can take many forms, such as correlation between ratings or cosine of those rating vectors. As in user-user systems, similarity functions can use normalized ratings (correcting, for instance, for each user's average rating).

Second, the system executes a recommendation stage. It uses the most similar items to a user's already-rated items to generate a list of recommendations. Usually this calculation is a weighted sum or linear regression. This form of recommendation is analogous to "people who rate item X highly, like you, also tend to rate item Y highly, and you haven't rated item Y yet, so you should try it".