The first recommendation feature at Amazon.com, BookMatcher, did not work very well. It required 20+ ratings to make recommendations. The recommendations leaned heavily toward bestsellers. The recommendations were often uninteresting, spurious, or obvious. Adding insult to injury, the system was always in danger of falling down under load.
As another side project (see "Early Amazon: Inventory cache"), I started working on a replacement. I wanted the recommendation engine to be fast, scalable, and able to work from just a couple ratings or purchases.
This was not what I was supposed to be doing, mind you.
I never got a chance to work on the BookMatcher project. I wasn't supposed to be working on recommendations. There were other, official projects that consumed my normal waking hours.
This was something I did on the side. My twisted, geek idea of fun.
I hacked. I tested. I iterated. With time, I built something that worked well.
I showed it to a few other people. Right around then, there was a major redesign of the Amazon.com website. Before I knew it, I was rushing my little prototype out the door.
A talented engineering manager named Dwayne helped put a spiffy UI on top of the recommendation engine. We rolled it out with the other changes. The new feature was called "Instant Recommendations".
Bookmatcher and Instant Recommendations co-existed on the site for some time. Eventually, low traffic, performance problems, and poor recommendation quality caused BookMatcher to disappear.
Instant Recommendations expanded to take its place, and eventually formed the backbone for much of the personalization at Amazon.
Such a useful and valuable feature came from work I was never supposed to be doing. In there somewhere lies the secret to innovation.
Subscribe to:
Post Comments (Atom)
9 comments:
How was your recommandation system related to academic ones ?
So do you recommend Amazon adopt Google's 20% time program?
To an outsider, Instant Recommendations sounds an awful lot like how Findory works. You get results after just a few articles, and it doesn't require nightly batch jobs to recalc.
Does Findory use a similar approach, or is personalizing news and blogs that much different from personalizing book recommendations?
mb, the algorithms behind Findory are quite a bit different. Recommending news stories has some unusual challenges. In particular, news stories are generated much more rapidly and expire much more quickly than books or other products.
Anonymous, which academic recommendation systems?
Anonymous #2, I do like the motivation behind Google's 20% time program. I'm not sure if 20% time is ideal for Amazon, but it does seem to work well for Google.
After reading your post, I just read your paper "amazon.com recommendations" http://ieeexplore.ieee.org/xpl/abs_free.jsp?arNumber=1167344 . I guess that's the algorithm you are referring to in your post? In the end, it is somewhat similar to association rule mining (customers who bought A also bought B) except that you calculate the similarity of the items later on. What do you think about association rule mining and how does it compare to your algorithm? I missed that comparison in your paper...
With Amazon apparently testing the biggest site redesign in years, I'd love to get your thoughts on where you think Amazon still has opportunities to improve their game.
-Chris
Hi, Chris. I'd love to see Amazon do personalized search, different search results for different people based on their past behavior and interests.
Stepping back for a second and looking at the bigger picture, I think it is hard to say that Amazon is anywhere close to done. The experience of shopping at Amazon is hardly effortless, full of discovery, or even all that pleasant.
Going to Amazon should be like walking into your favorite store, the nearest shelves piled high with things you like, everything you don't need fading into the background. When you walk up to an item, everything you need to quickly evaluate it and decide whether to buy it should float to your attention. Buying should be effortless, a couple clicks at most, with no unpleasant surprises (such as hidden shipping charges, delays, or belated out of stock e-mails).
Amazon has taken some steps toward that vision, but is a long way from there.
this is great! when did the events you detail occur? in other words, when/in what year did Amazon introduce Instant Recommendations?
Being a Masters student doing thesis on Web Page Recommendation, I would like to know more on the recommendation algorithms used currently at Amazon and also the scope of integrating semantic web technologies with web page recommendation.
Post a Comment