You’ve probably noticed the creepy effect where you consider buying something online, or maybe just look at a page for something that happens to be for sale, and for weeks afterward you get ads on totally unrelated websites for that thing or similar things. The more reputable online ad brokerages offer a degree of control over this effect (e.g.: Google, Microsoft). This study investigates exactly what effect those settings have on the ads observed by automated browsing
agents. The basic idea is to set some of the knobs, visit websites that will tell the ad provider something about the simulated customer’s preferences, possibly adjust the knobs again, and finally record what is being advertised on a general-interest website.
A great deal of the paper is devoted to statistical methodology. Because the ad provider is a stateful
that advertiser has exhausted their budget for the month), it’s vital to avoid as many statistical assumptions as possible. They use permutation tests and supervised classification (logistic regression), both of which make minimal assumptions. They’re also very careful about drawing conclusions from their results. I’m not much of a statistician, but it all sounds carefully thought out and plausible, with one exception: heavy reliance on  to the point where some journals no longer accept its use at all . This is exactly the sort of research where p-values can mislead; if I were reviewing this prior to publication I would have encouraged the authors to think about how they could present the same conclusions without using significance testing.
Now, the actual conclusions. Only Google’s ads were tested. (Expanding the tests to additional ad brokers is listed as future work.) They confirm that turning a particular topic (dating) off in the preferences does cause those ads to go away. They observe that two highly sensitive topics (substance abuse, disability) that do trigger targeted ads are not controllable via the preferences; in fact, they are completely invisible on that screen. And the most interesting case is when they set the ad preferences to explicitly reveal a gender (man or woman) then browsed a bunch of sites related to job searching. Simulated customers who claimed to be men got ads for a
career coaching service which promised better odds of being hired into
$200K+ executive positions; those who claimed to be women did not see these ads.
This last example clearly reflects the well-known
… We consider it more likely that Google has lost control over its massive, automated advertising system. Even without advertisers placing inappropriate bids, large-scale machine learning can behave in unexpected ways.
There’s a lesson here for all the
big data companies: the best an
unbiased machine learning system can hope to do is produce an accurate reflection of the training set—including whatever biases are in there. If you want to avoid reduplicating all the systemic biases of the offline world, you will have to write code to that effect.