Just a few points. First, there is only one website where we gather the user behavior. And yes, we predict interests for people that are clients AND that we can identify (either by logging or other means). This identification allows us to make the link between online data (web profile) and offline data (CRM).

For us, the hardest part is to aggregate the web raw data from the web logs (several Gb per day) in a web profile for each user.

When confronted to unidentified visitors, we find interest groups after they have visited 1 or 2 pages on the websites using the human-defined rules. But of course, there may be other means of doing.

We have the following levels: URL -> Categories -> Interest Groups. Even if categories may have hierarchies, we don’t explicitely use this concept in our solution for the moment.

I’m planning to write more about these topics in the following weeks

Interesting project. If I understand you correct, then you build a flat static model based on features independent of the behavior on the website . The target variable then is deduced dependent on the sites visited, so that you can predict the interest of other visitors as long as they a) are logged in and b) are already clients. As you said, they do not even have to visit the site to get scored.

So far, this strategy sounds good to me

I guess the hardest part is to calculate the label. Merging visited sites, time spent on page and maybe the multiple categories of a site into one crisp value is, well, hard.

Even harder is to identify how interested a visitor, who is not logged in, for certain categories. Now we are leaving the area of flat models, now it is getting interesting …Any plans in this direction ?

Another point is, that the categories may have a hierarchy (do they ?). I recently experienced that building models for hierarchical models is really non trivial. Any thoughts on this issue ?

kind regards,

Steffen

PS: More posts like that. I like them technical