Hacker News Reader

Comment by jtraffic | Hacker News Reader

> At some point, a human is going to take a look.

I recently saw a talk by Foster Provost, a big ML guy at NYU. The main points of his talk were that using fine-grained behavioral data (like browsing history) is better than demographics (at least in his context: predicting ad lift), and he proposed a way to interpret the model (somewhat). I left feeling disappointed. His system for interpretation was super post-hoc and tenuous, IMHO. It felt like a computer scientist doing social science (because it was).

I think ML is great for lots of things, but there are still lots of problems with using it in systems with humans.

An example: if you do a Google search for "Amazon <some book>" then you'll almost always get an ad from Amazon. So Amazon pays for a click that would have happened anyway. Maybe Amazon does this to crowd out other advertisers, but maybe it's just the algorithm being stupid and nobody is watching.

View on HN

It's not being stupid it's designed that way. "Gee that'd suck if somebody else bought ads for all your organic keywords. Sorry we don't have a better way to prevent that, you should bid on those."

Even if Google did provide a provision for blocking ads for certain keywords (and they do have one), who controls that system? Amazon is also a bad example since many 3rd parties affiliated with - and driving traffic to - Amazon buy those keywords.

Each individual advertiser on Google can provide "negative keywords," which are keywords that the advertiser would like to make sure they never show for.

Yes, you're right that buying one's own brand keywords is a defensive move. The idea that Google would show a relevant ad for a competing e-commerce site when a user is searching for "Amazon" isn't scandalous, it's the business model.

I think that's entering an area of ad ethics. Yet, its an ethical violation that companies like Google make today. If I search for "Wealthfront", I get an ad for Wealthfront, Betterment, and TDAmeritrade, in that order, before an organic result for Wealthfront. If Wealthfront didn't pay up, even though it is _obvious_ what I am looking for, the result I am looking would be buried. That's a protection racket.

Rank-based advertising, like in search results, is fundamentally unethical. To do it ethically would require perfect comprehension of what the searcher is seeking, so that you can prioritize organic, helpful results. Google never prioritizes organic results.

What? It's not a protection racket. Google is a private enterprise who is free to show whatever content they like on their page. Not showing your site is not illegal, nor is it unethical. A protection racket involves some illegal compulsion to pay for the protection. That's what makes it illegal.

It's not a protection racket but it is close to becoming a monopoly, which is also illegal in some circumstances.

This sort of anti-competitive coercion through overwhelming market power is why monopoly laws exist in the first place.

I don't think you understand what a monopoly entails. Having a large marketshare is simply not the only qualifier. It's also not remotely anti-competitive. That's just not what that word means.

If I search for "Wealthfront", how likely is it that I don't know they exist and need help finding their homepage, and how likely is it that I want to read about the type of service they offer?

> An example: if you do a Google search for "Amazon <some book>" then you'll almost always get an ad from Amazon. So Amazon pays for a click that would have happened anyway

This is absolutely understood, and in fact deliberate. Amazon does not want another advertiser sitting above their top position in the organics. It weakens mind share, and risks losing a click.

Ok, admittedly my example was bad. However, the replies are all basically assurances of the form: "Don't worry, a human at Amazon made that decision, so it's ok."

I'll repeat that ML is useful, and I'll add that it's useful in human systems, but its utility is constrained because for the really important stuff you eventually have to have a person check its work.

In my experience it really depends on the vertical. For some things (ford cars) demographics don't work well, but for others (mercedes) demographics like household income and age perform really well.

I think their is so much input for rtb that it is prime for ML. Most exchanges have some basic ml auto-optimization, but eventually someone will automate almost the whole thing.

Reply on news.ycombinator.com