Thu 02 Feb 2006 10:27:10 PM PST
John Buckman has an interesting problem: generating fair "customers who bought this also bought" lists.
I think the solution is not to focus on a shorter time span, but to divide by total sales.
Let's say everyone who comes to the bookstore gets Linux Journal and Martha Stewart Living.
Half of the customers also buy Confessions of an Advertising Man and The Cluetrain Manifesto, The other half buy The C Programming Language and Linux Application Development.
Someone walks in, gets a copy of The C Programming Language and checks the "Customers who bought this also bought" machine. It's going to show all three of the other products the same, even though two are just plain popular, and one is strongly correlated with the chosen book. Divide by total sales, and you get a stronger recommendation—twice as high for Linux Application Development—not just a recommendation for what's popular in general.