Getting a lot of email questions from the last post. . . so I dug around try to figure it out. . . found this but didnt help too much so I asked around for a no bullshit answer. This is basically what I gathered, there are two kinds of behavioral ad targeting
1) Bullshit Behavioral - Basically what Claria is doing, serving ads to you when you are surfing another website. Using the content/category of another site you are on to serve you relevant ads. Some times this might include serving you travel ads two or three pages after you left Expedia. More sophisticated ones combine a few websites and do some calculationg around what that means as far as your likelihood to click on an ad (silicon.com + travelocity = United Ticket to SJC). No even a simple implementation of a logit model, mostly hardcoded.
2) Wanna Be Amazon Behavioral - Think of the way Amazon serves you up product recommendation, but instead serve up banner or text ads instead. Using (People who bought product X also bought product Y) and applying it for ads (People who clicked on ad X also clicked on ad Y).
So there you go, I’m sure other people have their own secret sauce they dont want to share with me, but at this point doesnt seem like such a black box, rocket science, or breakthrough at all.
These things might sound simple but are very hard to implement. Depending on the interface, the whole trick is to do this reliably enough, so that everything is relevant. What would you consider something more interesting?
Comment by Joe — September 22, 2005 @ 9:23 pm
Not saying its not effective, the data shows it is. So it might not be very “interesting” intellectually, its “important” from a business perspective. Amazon itself admits that its recommendation algorithm for collaborative filtering is not too sophisticated but the incremental processing cost (buy more machines) of a more sophisticated implementation using k-mean clustering (or some other method) overwhelms the profit improvement. So I get your point that the hard part is not the algorithm but the operations and implementation when the matrix grows to millions and millions of row & columns.
Comment by Administrator — September 23, 2005 @ 10:25 am
Put a different way, the challenge is in amassing the data that can underlie such targeting, rather than the analysis of the data itself.
Comment by Ray — September 23, 2005 @ 12:38 pm