The Network Behind Amazon Recommendations
It’s no secret to anybody that the websites we use every day have tons of data about us. We explicitly provide sites with all kinds of data about our lives when we fill out Facebook or LinkdIn profiles, and we implicitly provide them with data about what we like or dislike when we watch something on YouTube or search for something on Google. And with all this data at their disposal, its no surprise that companies take advantage of it to get you to use their service more and ultimately increase their profit. Now with something like Facebook that’s literally a “social network,” it’s not hard to see how the data relates to the networks we’ve been discussing. But what about something a little more subtle, like Amazon’s recommendations?
How does Amazon choose what products to recommend to you? Of course, they can recommend things related to what you’ve bought and/or rated highly in the past, but what exactly does “related” mean? And furthermore, sometimes buying one thing can lead to a recommendation for something that seems completely out of the blue. Where does that come from?
The answer is that Amazon doesn’t just look at you in isolation; it uses some of the massive amounts of data they have on other customers too. Amazon might not have any data about who your friends are, but it has something even better: it knows who’s bought the things you’ve bought and liked the things you’ve liked, as well as all the other things that they’ve bought and liked that you might not have even seen or heard of before. And although it might not be immediately obvious, this is a network. Imagine a graph where every node is an Amazon customer, and there’s a node connecting two people if they’ve bought the same product (maybe weighted in some way if they also gave it a rating). This enormous graph has large clusters of people who bought the same thing. If Amazon can determine that you have a strong connection to lots of people within a cluster for a product you bought (i.e. you have multiple edges connecting you) and that those people have strong connections with lots of people in a cluster for another product, then it only makes sense that Amazon recommends you buy that product, join that cluster, and restore some triadic closure (as well as giving Amazon more money).
And although Amazon doesn’t say much about how effective the recommendation system has been for them, they definitely seem to have a lot of confidence in it, as they recently patented “anticipatory shipping.” Essentially, they’re planning to ship products that haven’t been ordered to areas where they believe they will soon be ordered, thus reducing delivery time. Note that this is probably more likely to be used for products that you have a history of buying regularly, rather than shipping new products you’ve never bought before, so it isn’t quite the same as the regular recommendation system, but it still says a lot about how much confidence Amazon has in their ability to use its data to predict what you’re going to buy.
http://fortune.com/2012/07/30/amazons-recommendation-secret/