Caterpillar Cowboy

heyitsnoah:

Partly because I haven’t written anything of any length in awhile and partly because I’ve been thinking about a bunch of different stuff lately, I’ve got an entry chock full of random thoughts.

So,…

RE: Recommendations (I used to be the Technical Product Manager for Amazon’s recommendations group)

The trouble with charging for recommendations is when you’re wrong. When something is free, your tolerance for mistakes is higher. If you’re paying, you expect results.

Common pitfalls:

• The Harry Potter effect: Some products are so popular, anyone who has bought anything has probably also bought Harry Potter, making it seem to the algorithm like a good recommendation but in effect it’s not relevant at all (or at least, not serendipitous).
• Special edition effect: Because you bought X and Y, you also bought Z. Other people who bought X and Y, also liked Z but bought Z in red, or XL, or widescreen (call it Z prime). So you get recommended Z prime, even though you already own Z. Whoops. This is a data quality problem that is solvable, but difficult if you have a catalog of Amazon’s size.
• The WTF effect: When you base recommendations off of a fairly obscure product, you don’t have a lot of data to work with. As a result, sometimes the relationships the algorithm comes up with don’t make any sense at all, and are simply a result of noise derived from small sample sizes. People laugh / critique Amazon about this all the time: “These two things are obviously different. How stupid of them!”

People would probably say Amazon’s recommendations technology is the best in the world, but even so it is wrong a lot. Thankfully (for them), they make money every time they are right and you purchase an item. And yes, that does make Amazon a fair amount of pocket change.