One big clump of recommendations engines...
I'm doing a bit of work around recommendations and recommendations engines at the moment, and I'm finding it really illuminating. The thing I think I'm most surprised by is how unclear the boundaries that surround the whole concept actually are - how everything seems to bleed into surrounding areas - where structure and categorisation bleeds into navigation bleeds into contextualisation bleeds into associations between things which bleeds into tracked user-behaviour in aggregate which bleeds into individual user behaviour patterns. It's all very... bleedy...
Anyway, so I'm thinking about all this stuff and I start looking at Amazon - and I know I shouldn't really be surprised but I suddenly get slightly overwhelmed by it. It's basically just loads of recommendations engines joined together with a tiny fragment of 'buy a product' goo. There's the page that greets you when you arrive (assuming you're logged in), above which you can click on a page to get specialised recommendations based upon purchases. On every single item you go to there's a recommendations aspect (people who bought this also bought...) - and then if you add it to your basket, then you get a recommendations page that shows you other things you might like to buy. The same thing happens when you add it to your wishlist. And then there's the recommendations engine that tracks you around the site, keeps track of every item you've looked at and works out descriptions based upon those - that's called "The Page You Made". And then there's the "New for You" page - a set of (you guessed it) recommendations based upon what's recently been published or released.Then there's the button that you click to "See more items like the ones in your wishlist". And then there's "Your Store"... I wouldn't be surprised if they tweaked your recommendations depending on where you lived as well.
All of which only makes it worse that seem to think they think I'm obsessed with low-grade sk8ter rawk...
Comments
Please stay on-topic, informative and polite. I reserve the right to remove comments for whatever vague capricious reasons seem reasonable at the time.
ewwww sk8er rawk is the exact product of corporate marketing gone really successful. fakes, phonies, an image that is precieved as tough and cool because of the clothing brands you wear, the "punk" music you listen to, and because you hold a skateboard in your right or left hand. sick really.
→ Posted by: kavi at September 8, 2003 2:51 AM
Coincidentally, Erik Benson wrote a piece on Amazon Recommendations - http://www.erikbenson.com/index.cgi?node=fortune-telling - in which he says "The goal of personalization is to get to know you so well that all it takes is a complex algorithm to predict what you will do (usually buy) next. We are trying to outspeed the computation power of your brain" ... and ... "We don't care why you bought something, because whatever the reason was, some or all of that reason may apply to another person and they may end up doing the same thing that you did if they've done things similar to you before." ... which is all good stuff (doesn't mention the feedback loops which feel a bit like time-travel: if you hadn't been recommended the book, would you still have bought it?) BUT I think misses one of the key features of Amazon :: the recommendations are used to establish context. They perform the same function as: the reviews on the back of a book saying "The New Mr X!"; that some books on the shelves face outwards and others don't; similarities of cover art designating style/subgenre. All of these things can be done in a bookshop because you see many books at once -- Amazon has to do it deliberately. So it's important to have in the "Customers who also bought" mix books you've bought already? I reckon so. Oh yeah, and back on topic: Not only is Amazon is bucket of recommendations systems, it's also a great fulfilment company. That's all you need.
→ Posted by: mattw at September 8, 2003 8:30 AM
It would be really good to see even a tentative categorisation, if you come up with one. I've had my interest in reputation systems stimulated by the way the Elance one has broken; it's set me thinking that very, very few people seem to actually understand them.
→ Posted by: Ian at September 8, 2003 9:21 AM
I sometimes have to look artist information up in Amazon for work (I'm an online music journalist). This then becomes part of my recommedations! Amazon currently thinks I like The Coral, Blazin Squad and Bergman movies. This is not true.
I find the "People who bought this..." links quite useful and they have often led me to unexpected but rewarding book and music purchases.
→ Posted by: Garret Keogh at September 8, 2003 12:18 PM
From my own experience, the Amazon recommendations work better the more you use the site - the more products you browse, buy, etc. the more refined their computations. Or something.
Of course, having said this, I take much more stock in the 'People who bought this have also bought:' links, since these are filtered by individuals and their actions, not statistical guesswork.
→ Posted by: MacDara at September 10, 2003 7:38 PM
Tom, I had a slightly different ponder about recommendation systems a while back. I just wondered whether search engines could be improved by some form of reputation management, and the way I considered this working was with a network of recommendations. It's basically an idea to overcome the fact that search engines at present are fine for precision and recall, but don't help you to judge the "trustworthiness" of information. Any thoughts (beyond the obvious personal data protection issues)?
→ Posted by: Mark Thristan at September 11, 2003 11:36 AM
Well of course in one way they are. Google's Pagerank (as my good friend Mr Webb suggested a while back) nothing but a recommendations mechanism for web-pages. It's reputation component is in the number of people who think a page is interesting or useful enough to link to. The only difference is that there is a difference in register. With something like Amazon, the fact that someone else has bought two objects has created a link between those objects of a greater or lesser strength. The two objects, however, are of the same register. They are both 'things you buy'. In Google, there are things you look for and things you link to - although I have no doubt (since we can see it in their adverts on the right) that the fact that you choose to click on one link rather than another is recorded and reused in some way... A variety of other people have been trying to talk about positive and negative links recently - some of whom as a result of reading one of my old posts - On the responsibility of linkage - but others working completely independently. The gist of their work seems to be that there should be (at least) two kinds of links - one that is a recommendation (as all links are today - albeit tacitly - in that they build traffic for the person linked to) and one that is an obvious and clear statement of disagreement which would be something that a search engine could parse and utilise. I'm not sure I know how useful or functional such a negative link would be - nor do i know how you'd implement such a thing or how prone to abuse it would be - but it's certainly interesting as a concept.
→ Posted by: Tom Coates at September 11, 2003 9:49 PM
Could I just say to "south beach diet" who keeps posting spam comments on this site that I'm just going to continue to delete them as fast as you post them, add them to a blacklist and circulate them around the community. At a certain point I may start sending you bills for the advertising on my site that you are undertaken, working on the basis that you are implicitly agreeing to this site's Terms and Conditions of use.
→ Posted by: Tom Coates at February 21, 2004 12:44 PM