WSJ Article talks about determining gendering after 5-10 clicks

Article didn't address how, but I'm interested in how this is done

[Jeremy_H]

I was reading a Wall Street Journal article in my newspaper today by Jessica E. Vascellaro. The title was called "Online retailers watching your every click".

[post-gazette.com...]

The article is interesting, addressing how websites are dynamically changing their website to better entice potential customers.

The article touched on how websites were able to gain certain metrics such as the users time based off the timezone associated with the user's IP address.

The article states:

"The company says it can typically determine a shopper's gender after about five to 10 clicks."

This is new and very interesting to me. Obviously the audience of the article was to general consumers, and not programmers, so she didn't address the technical aspects or theory, but I'm very interested in what sites use to determine gender based on clicks?

Is it well regarded males click on certain things and females click on other things? (Irregardless of age, income or nationality?)

If you have tricks to determine gender, I'm very interested in them.

Thanks

[gregbo]

I read some time ago that Microsoft was doing some research in determining the gender of a surfer based on clickstream data. Like anything else on the web, this is highly susceptible to fraud because clickstreams can be easily generated to fit given patterns.

[cgrantski]

Classic (well classic within the last 10 years) predictive modeling of behaviors. Cool stuff, I used to do it in graduate school and wish I had found a way to keep on that career track without sinking into the marketing industry swamp. Anyway, they get click data on people for whom they already know the gender then run statistical programs that find patterns of clicks that correctly discriminate between males and females, then they apply what they learn to fresh data where they don't know the gender, making the assumption that the same rules will still work. It's usually done for each situation ... in this case each site. There are a lot of techniques for extracting the rules and the rules can be so complicated that they might have to be expressed as a mathematical model rather than a simple checklist. It really wasn't possible to do this before computers.

Guaranteed they are not talking about generic, generalizable, cut and dried signs and signals of gender.

[whoisgregg]

Basically all they need to do is have multiple links to each page with varying anchor text or images and track which link the visitor uses. If you have registered users who provided gender as part of their profile, you track their clicks as the control group. Patterns of "click tendency" in that control group can be used to make a pretty good guess as to gender (and other characteristics) of unregistered visitors.

Even if you just have two variations of links to each page and it's a small differential between the genders you can become reasonably certain if as little as five clicks. If the first five links versions that a new visitor clicks are the preferred link versions of 55% of the registered women and 45% of the registered men, then the statistical probability that the new visitor is a woman is something surprisingly high, like over 90%.

Of course, it won't work out that way so you might need more clicks to be sure. Even then, some visitors may be "gender neutral" over time and sometimes the guess will be wrong but that extra bit of targeting, even if it's only right some of the time, can still be incredibly worthwhile.

It really wasn't possible to do this before computers.

True, but before computers the number of web site visitors was so low as to not make the attempt worthwhile. ;)

[plumsauce]

but before computers the number of web site visitors was so low

classic!