For the March 2012 issue, I wrote about a new subject for me, and one that I hadn’t really seen much discussion of elsewhere: recommendation algorithms.
Whenever you go to Amazon to buy something, there on the front page is a set of items that you might want to purchase. These recommendations aren’t merely what a lot of people are buying (for example, I never get offered a Justin Beiber CD), but they do seem to be specifically geared to your wants and interests. Yes, they can be hilariously wrong (for example, for December and January I seem to get a lot of references to kids’ toys because that’s the only real time we buy them, for our nephews and nieces and friends; children), but in reality I’ve discovered items in those recommendation lists I’d never heard of that I’ve subsequently bought.
Similarly, if you’ve used NetFlix, you’d see the same kinds of reasonably-accurate recommendations there too. How do they do it?
The answer is some specifically-tuned statistical algorithms known as recommendation or prediction algorithms. These range from the really simple ones (“you bought Abbey Road by the Beatles, I’ll recommend Revolver the next time I see you”), to Slope One predictors, to clustering algorithms, to collaborative filtering. No matter what is used, the intent of using recommendation algorithms is to improve your chances of getting a customer to buy something else from you (or, even more fundamentally, not to go elsewhere, to keep them on your site). The more sophisticated algorithms are measurably more effective at retaining customers or at getting them to buy, or they are evolutionarily discarded.
This article first appeared in issue 319, March 2012.
You can read the PDF here.
(I used to write a monthly column for PCPlus, a computer news-views-n-reviews magazine in the UK, which sadly is no longer published. The column was called Theory Workshop and appeared in the Make It section of the magazine. When I signed up, my editor and the magazine were gracious enough to allow me to reprint the articles here after say a year or so.)