“Tell me who your friends are and I will tell you who you are”
That saying cannot be more true today – especially with all the available social networks. How does Amazon make recommendations to its users? While they don’t tell us everything they do, Amazon does reveal one part of it at least when you login:
“These recommendations are based on items you own and more”
I don’t find recommendations based on items I own very interesting, I am more interested with the two words at the end – “and more”. What other ways are there to provide recommendations?
User demographics have been the most popular way to do this for the longest time now, and with the rise of social networks I believe this is about to change. What is the problem with using demographics for recommendations?
The biggest problem I see is demographics and statistics go directly against personalization by definition. A decade ago that was probably as close as one can get to personalization, but today is this still true? Today, everybody – including the neighbour’s cat – has a Facebook profile. All these social networks are sitting on a gold mine and they know it. Demographics are useful when the medium you are delivering the content through delivers content by the same demographics such as a TV. As a TV advertiser you need to know the demographics of the product you are advertising to know at what time of the day you can maximise your reach to that demographic. Unfortunately today’s web does not work that way – imagine Facebook where on weekdays between 10am and noon you can only add and talk to stay at home moms? or between 4pm and 6pm its kids between the ages of 8 and 15, or between 8pm and 11pm its men between 30 and 40.
Today’s social web is not divided by demographics, whether you are 15, 25, 35, 55, or 60 you can still get on Facebook and find content that interests you.
Consider the iTunes Genius feature that recommends songs? How does it do it? Obviously there is some complex algorithm that answers the question: “If user 1 purchased songs a,b and c, and many many many users purchased songs a , b and d, what is the probability that user 1 will like song d if song d is the same genre as songs a and c for user 1?” This is the bottom line of the majority of today’s recommendation engines. Based on my experience as a customer it comes down to:
- What did I already buy? - or listen, or view, or read, or…
- What did others buy? - or listen, or view, or read, or…
- How big is the intersection between our product selections? and is it enough to conclude that I may like what they bought?
So is this personalization? Isn’t this similar to how radio stations select songs?
For example, what I would like to see on Amazon is the recommendation engine would learn about me based on what I provide on the profile. By pointing Amazon to my Facebook, blog, twitter, linkedin, bit.ly bookmarks and more, Amazon would learn about what interests me based on my thoughts here, my rants on twitter and who I follow, my friends on Facebook, the events I attended and who from my friends went there, and so on.
Finally, if all this sounds like too much information to give away, although it is already publicly available, then you don’t need to use it. Think about it how neat would it be if Amazon was able to learn through my twitter that I started learning CakePHP and sends me a special offer for 50% off a CakePHP book? or that I started using twitter 3 weeks ago which coincided with when I started posting more often on here and recommend a book about promoting my blog via social media?