Reponse to "Amazoning the News"

The amazon/newspaper article is interesting, but it oversimplifies a design. However, I think some of the details are what make it interesting. What recommendation algorithm would be used? Amazone uses an item to item comparison to determine what else would be of interest. Essentially working through who bought each item, checking what else they bought, and returning common results. The reasoning behind that is that although it is processing intensive on the backend, it can return highly relevant recommendations without knowing much about each user.

The question is, what recommendation algorithm would media site use? Using an item to item comparison similar to Amazon's would run into a problem withe the relatively short time that a story is up-front. It wouldn't be feasible to wait for a nightly process to go through the stories that were also read by individuals.

A person to person algorithm might work, identifying similar users based on story choice. However, as Amazon found out this method is flawed, because to be a useful system a critical mass of readers are necessary. Unfortunately, once that critical mass is reached, comparing each reader against every other reader quickly becomes unbearably intensive.

Two other methods are feasible, but provide less valuable results. The first is a simple keyword match, that would pull up past stories that were based on a similar topic. This functionality is already built into many search engines. The second method would be profile creation, essentially marking each story with a category and associating that category with the user. For instance, if someone read 5 stories about the Denver Broncos, the system would assume that they would be interested in other stories related to the Broncos. Unfortunately, this can lead to embarrassing categorization...similar to the recent jokes about a Tivo assuming you're gay because you regularly record Queer as Folk and Will & Grace.