Illustration
collaborative filtering and recommender systems

The coloured boxes below show the performance of two different recommender systems that predict a customer's preference for movies they haven't yet seen.  In these boxes, successful predictions lie on the main diagonal, from bottom left to top right of the squares.

The left-hand diagram shows the predictions based on average preferences. A customer's "recommended" score for a movie is simply the average preference score for the movie. In other words, popular movies are recommended and unpopular ones aren't. The preponderance of 'cool' (blue) numbers, and the almost horizontal band, show that the predictions are rather poor, and this is because individual customers have different patterns of preferences.

In the right-hand diagram, preference patterns are extracted from other customer's data and an algorithm applied to predict whether a particular customer will like or dislike a particular movie. The 'hot' regions, and the diagonal band, show that the predicted preference scores and the customer's actual preference scores are much better matched.