Recommendation Engines

Was off my game (not sure you’d notice) Tuesday afternoon so wasn’t very clear. In relation to recommendation systems and things like FaceBook and Amazon, specificity here matters. They both have ways of recommending things but they are quite different in how they work, and why. FaceBook is about selling ads to advertisers, just like TV and newspapers. So when it puts ads in there it is using what it thinks it knows about you, based on what you and your friends have done on FaceBook in the past, combined with how much advertisers are willing to pay for their ad to someone that Facebook has defined as like you. So this is not a recommendation system, it is an advertising regime and the algorithmic systems that Facebook uses are about targeting advertisers to you. This makes it fundamentally different to the recommendation algorithms used by services such as Amazon and iTunes because for these latter services there is an enormous catalogue of material and as retailers they don’t care which one you buy, just that you buy. Facebook, on the other hand, since it sells ads which, as ads, want you to buy that thing rather than some other thing, has to care about what you choose and why – this is the entire premise of advertising. (In other words the book shop doesn’t care which book you buy, as long as you buy a book they stock, though once in the book shop some publishers will use different point of sale advertising to try to get you to buy that particlar book. Amazon and iTunes music store are like the book shop, they just want you in the shop, Facebook is more like the publisher, they not only want you in the shop, but then they want to sell ads to those in the shop.) Advertising is not and cannot be driven by recommendation algorithms because, for Amazon and iTunes, these are anonymous peer driven (anonymous because you don’t know them, peer because funnily enough there are other people who seem to like things you do, and on that basis there’s a pretty good chance you’ll like what they like too.)

Amazon on the other hand uses its data, harvested from what people buy on amazon, to data mine it to build its recommendation system. This is not a lot more complicated than using what I buy to define my profile, and then matching it with similar profiles. Once this is done it is possible to make suggestions based on what other people who appear to be interested in the things I’m interested in are interested in (there’s a lot of interest there). However, it is not trying to sell me anything specific, it just wants to help me find new things, in particular things I might not have noticed, on the reasonable assumption that I’ve bought there before, so am likely to do so again. In other words things don’t appear because they’ve been paid to be there (which is Facebook), they’re there because lots of people who buy those books also buy these ones. Because of Amazon’s scale (how many sales it makes) it has an enormous amount of information from which it can build its recommendations. It also lets you rank and rate its recommendations, which is handy and of course lets their algorithm became better. iTunes recommendations work the same way, it is simply using sales information so that people who like Bonnie Prince Billy are also likely to enjoy Bill Callahan. What is of value here is that it usually only takes about two clicks to find stuff that you often don’t know, and you can then decide if you’d like to listen to it. So, in relation to producing recommendation hierarchies it is quite resilient.