I have no inside knowledge of the actual algorithm. But I always assumed it was just a simple query of all the customers who bought that book, then aggregate all the book purchases by those customers. I doubt there is any specific filtering on purchases made at the same time.

Thanks for your reply. so according to you, they will find all customers who bought that book, then find other books that those customers bought and suggest it to the user.

It makes sense. So they are filtering the suggestions by category. For instance if another customer who bought that book and other books and also some other items, then you are only interested in the other books purchased by that user and discard the other items.

But I've also seen that when you are browsing for a TV - in that section of "customers who bought this also bought" - it will list other items like DVD players, home theater systems, remote controls and even cables - not just TV's.

Well if u think abt it - all those items do fall under the parent category of electronics - but not in the sub category of TV. But what is interesting to note is that all these items are related to a TV-they are not showing stuff like digital cameras or camcorders which also fall under the parent category of electronics.

Thanks for your opinion though.

I'm sure they do some sort of category-based filtering. Or, unrelated items fall out just from "popularity" numbers. (For example, if 500 people bought that TV, and those 500 people colelctively bought 300 DVD players, 400 DVDs, 100 VCRs, and 1 Tivo, the Tivo probably wouldn't make the "also bought" list.)

category-based filtering

Not always. I recently bought a "Health & Personal Care" item that had a few books shown in the "customers also bought" section. The books shown were a perfect accompaniment, so I ended up buying one of them. :)

OK, "related category-based filtering" :)

Collaborative filtering is prob what they are using
[en.wikipedia.org...]

Great link, thanks bateman_ap! :)

Why do you need a category filter?

If the query is done based on items in an order, and not really everything a customer has ever purchased, the items will have a tendency to filter themselves, sometimes better than an algorithm could probably figure out.

I've done similar things myself. When the sample of data is large enough, the statistical trends can be quite accurate.

Greg Linden designed and developed the recommendation algorithm at Amazon.

Forget Wikipedia try Amazon.com Recommendations: Item-to-Item Collaborative Filtering, Greg Linden, Brent Smith, and Jeremy York [computer.org].

- George

Thanks for all your responses guys.

George - thanks for the link - thats very helpful information. Thanks for sharing.