Ad click prediction: what are the negative examples?

Question

I am analysing the log of a website and I would like to build a classifier to predict the users that are likely to click on an Ad.

The Ad can be displayed to the visitor several times.

To build any classifier I need positive and negative examples:

• The positives are the visitors who clicked on the Ad (easy).
• The negatives are the visitors who saw the Ad but didn't click (not very obvious).

Question

• Is there a convention about how/when to consider a user as a negative example?

I presume that I should define a threshold of impressions (views) per user, if the user reaches this threshold without clicking on the Ad, I consider him/her as as negative example?

Any reference or guidance is highly appreciated.

Answer 1

You're overthinking it. You might not need a threshold. Start with the simplest approach you possibly can: If you showed the ad to a visitor, then that's a negative example. Each time you show an ad to a visitor, you end up with an instance, whether positive or negative. If you've showed the ad three times to the same visitor, you end up with three examples. (Maybe all negative; or maybe one is positive and two are negative; but that's fine.)

Bonus tip: Do research methods for handling class imbalance.