WebSite user behavior modelling

[playcat]

I want to model my site behavior, finding out what would be a best way to show certain ads. Now, there is a lot of ways to show them: I could, for example, always show the ads I earn the most from, but that doesn't make sense because if user doesn't click once on an ad, he might not click again . On the other way, if I go and show them ordered by price, I might earn "some" cash, right?

There are lot of ways to do so. Hence my modelling need.

 

Can I model user behavior using database with user clicks? Should I use that db to create some sort of distribution for modeling users?

 

I don't have much experience in modelling, only one semester at my uni.

 

Thanks in advance

[the tree]

I suppose you could model user behavoir, clicking from one page to another with a rather large Markov chain, although I'm not sure how the data from that would help optimise the choice of adverts.

 

Another approach that I'd propose would be:

• Assume that you only need to 'try' an advert once, that is that once a user has seen it and not clicked, they aren't going to click it that session.
  
• Find the probability of an advert being clicked at all (the amount of times the advert has been clicked divided by the amount of unique users exposed to it).
  
• Multiply that probability by the profit from a click, to get the expected profit per user for each advert. (this should be a very, very, small number).
  
• Order the adverts from highest expected profit to lowest.
  
• Either by some complicated click counting with cookies and whatnot, or by just by depth from the home page - expose your users first to the most potentially profitable adds, then once those haven't gained a click move on to trying the others.

[Dave]

The algorithm above sounds pretty good but still (potentially) has the issue of adverts which never (or virtually) never getting clicked failing to appear. Perhaps instead of ordering them according to expectation you might try picking one with a probability relating to their probability of being clicked. If you choose this so that each advert has a non-zero probability, for instance choosing the max of say 0.01 and their calculated probability, then you might get a better algorithm. But it's all speculation really, you need to try it out