3 ms·

> the predictions being made with it aren’t terribly impressive So what? They still have the data and can refine their methods tomorrow. Today their prediction

by pdkl95 8y ago

> the predictions being made with it aren’t terribly impressive So what? They still have the data and can refine their methods tomorrow. Today their predictions might be low quality, but they can retry as many times as they want. The problem is not the predictions they are making today; it's the many predictions (or inferences) are able to keep making in the future. > political groups Remember that some types of advertising is not targeted. Some political advertising or branding advertising is intended to reach "all voters"j or maybe a very broad category like "Everyone Californian of voting age". Branding campaigns don't care if you're interested in e.g. vacuum cleaners. They just want you to think of their name first every time you happen to think of or hear about vacuum cleaners. edit: (Multiple contradicting groups could be pushing ads at your (very general) demographic.) > science....or dogs Many scientists like dogs? > Amazon thinks that my interests No, they think that showing you textbooks and vacuum cleaners has a greater chance of increasing their revenue, according to various statistical models. Targeted advertising isn't about targeting what your are interested in. It's about letting other people target your with what they think they can sell you. edit2: Of course, it could also be a terrible model trying to use data in stupid ways. I'm just suggesting that there are many plausible explanations.

mattkrause 8y ago
Sure; It's obviously a mix of clever and stupid. One of the more memorable political groups was the Sandinistas. I have never been to Nicaragua, so....limited relevance. I actually just looked (increasing my likelihood to be targeted again?), and was surprised to learn that they're still around; I had no idea they out-lasted Iran-Contra. As for Amazon, I agree that P(bought a vacuum) is a good signal for P(will eventually buy another vacuum). Still, if they have this detailed psychographic model of me, you'd think they could do slightly better than what is essentially a zeroth-order prediction.
- pdkl95 8y ago
  > I have never been to Nicaragua, so....limited relevance Ok, that's probably some sort of error. > you'd think they could do slightly better It depends on how they made their model. With machine learning, it's easy to build something that does improve the whatever high-level metric they were tying to optimize (like "revenue") , but gives essentially random results for many individuals. This is especially true for any anything that wasn't fully covered in the training data/; a handwritten digit recognizer may work wonderfully on average, while giving random results random output when presented with anything that wasn't covered by the MNIST dataabsse. Sometimes this can be especially fragile[1]! [1] https://medium.com/@ageitgey/machine-learning-is-fun-part-8-how-to-intentionally-trick-neural-networks-b55da32b7196 https://medium.com/@ageitgey/machine-learning-is-fun-part-8-...

by pdkl95 8y ago

mattkrause 8y ago

Sure; It's obviously a mix of clever and stupid. One of the more memorable political groups was the Sandinistas. I have never been to Nicaragua, so....limited relevance. I actually just looked (increasing my likelihood to be targeted again?), and was surprised to learn that they're still around; I had no idea they out-lasted Iran-Contra. As for Amazon, I agree that P(bought a vacuum) is a good signal for P(will eventually buy another vacuum). Still, if they have this detailed psychographic model of me, you'd think they could do slightly better than what is essentially a zeroth-order prediction.

pdkl95 8y ago
> I have never been to Nicaragua, so....limited relevance Ok, that's probably some sort of error. > you'd think they could do slightly better It depends on how they made their model. With machine learning, it's easy to build something that does improve the whatever high-level metric they were tying to optimize (like "revenue") , but gives essentially random results for many individuals. This is especially true for any anything that wasn't fully covered in the training data/; a handwritten digit recognizer may work wonderfully on average, while giving random results random output when presented with anything that wasn't covered by the MNIST dataabsse. Sometimes this can be especially fragile[1]! [1] https://medium.com/@ageitgey/machine-learning-is-fun-part-8-how-to-intentionally-trick-neural-networks-b55da32b7196 https://medium.com/@ageitgey/machine-learning-is-fun-part-8-...