However, I like your conclusion that users start calling ads information as soon as it becomes relevant to them.

Explicitly asking someone whether or not they like the ad, defeats the purpose of the ad in and of itself. Rather, advertisers should be asking whether users like their products and what they could do to improve it.

Your example is a bit of a straw man as doing a 60/40 split on men versus women ads to random visitors is horribly stupid… when your rule learning method tells you to flip a coin then it’s FAIL.

Here’s an example that works: Learning that there is a correlation to men buying beer and diapers together. One only needs to learn that correlation over a representative sample of transactions.. consuming all of the data is unnecessary unless you are looking for the low frequency correlations.
If those infrequent correlations matter, then you can still sample if you do it smartly.

Second example: learning that your male viewers visit different pages than the women or enter via a different route, then you’d be able to target ads to them correctly. If page-type X – then assume men and show mens ads. You can learn that rule with sampling and smallish tests if done correctly.

The devil’s in the details.