In class we discussed different aggregation techniques and also were showed Arrow’s Impossibility Theorem. Even though it is impossible to produce a perfect ranking system, companies such as Amazon.com work toward creating the best recommendation system they can. When glancing through Amazon.com, one can see different recommendation systems at work including “Frequently Bought Together” and “Customers Who Bought This Item Also Bought.” The purpose of these recommendation systems is to drive users into the Long Tail. Anderson relates the need for recommendation systems which an example about Rhapsody in which, “the front screen of Rhapsody features Britney Spears, unsurprisingly. Next to the listings of her work is a box of ‘similar artists.’ … in three clicks, Rhapsody may have enticed a Britney Spears fan to try an album that can hardly be found in a record store.” Online recommendation systems make it possible for customers to easily find out of date or unpopular items with much more ease than browsing arbitrarily.

Online recommendation systems are essential to business models built off the concept of the long tail. Companies using this model have almost limitless capacity for merchandise whether it be books, music downloads, or DVD’s. However the drawback to having all this merchandise is customers simply browsing have to sift through massive amounts of unwanted merchandise before finding something that catches their eye. Without proper direction, most merchandise that lies in the tail of the graph will not be sold. If items are not selling, then companies are forced to pull them from inventory truncating the long tail further with each item pulled. If this process happens over a period of time repeatedly, the company will be back to the “hits economy” discussed by Anderson in his blog titled The Long Tail.

Various algorithms exist for online recommendation systems including the cluster model, search-based methods, and nearest-neighbor (Linden 2003). Amazon.com because of its sheer size uses its own algorithm called item-to-item collaborative filtering. The basic idea behind item-to-item collaborative filtering is to match “each of the user’s purchased and rated items to similar items, then combines those similar items into a recommendation list” (Linden 2003). One important thing to note is that since Amazon.com has so much inventory and customers, it is imperative that the algorithm run in real time. To insure this happens, Amazon.com stores a similarity matrix for all its items. The main steps of this process are as follows: identify items known to be of interest to user, retrieve similar items list, combine similar items list if multiple lists, sort resulting list from highest-to-lowest score, filter sorted list to generate recommendations list, recommend top M items from recommendations list (Linden 2001). Because the similar items list has already been precalcuated, this algorithm is able to run very quickly.

Most of this is from my paper pertaining to the Long Tail and online recommendation systems(mainly Amazon.com)

http://www.google.com/patents?hl=en&lr=&vid=USPAT6266649&id=NIAIAAAAEBAJ&oi=fnd&dq=amazon+recommendations
http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=1167344
http://www.wired.com/wired/archive/12.10/tail_pr.html