Amazon: similar items instead of similar customers
In the January-February 2003 issue of IEEE Internet Computing, Greg Linden, Brent Smith and Jeremy York of Amazon.com described item-to-item collaborative filtering, with which the store personalises its pages for each customer. The table of similar items is computed offline, and the online part depends only on how many items the customer has bought or rated, not on the size of the customer base or the catalogue.
Why it matters
Collaborative filtering moved from comparing people to comparing items, which removed the scaling limit the paper names for the traditional approach: with tens of millions of customers and millions of items it is impractical. It is the first description in the atlas of a recommender system by the people who deployed it.
The paper: Amazon.com has 'more than 29 million customers and several million catalog items', and recommendations must come back in no more than half a second. Similarity is the cosine between the vectors of an item's buyers; building the table is O(N²M) in the worst case and closer to O(NM) in practice. Click-through and conversion are said to vastly exceed untargeted content, with no figure. US patent 6,266,649 (Linden, Jacobi, Benson, 2001) is cited; the patent itself was not opened. Amazon's annual report for 2002 (10-K, 19 February 2003) lists pages tailored to a customer's preferences, including recommendations, among the site's key features, with net sales of $3.93 billion; it gives no count of customers, so the 29 million is the paper's own. IEEE Xplore's metadata give all three authors as University of Washington; the paper is signed Amazon.com. The record does not claim what share of sales recommendations bring: no document read gives one.