This is exactly kind of "if all you have is a hammer, everything looks like a nail" that turns me off about AI/ML community.
> don’t need to keep seeing hotel offers for Florida
This is not for the lack of smartness in code - it is that the conversion funnel is so narrow that "saving that $30 per month, by not showing you the ad, on a $1m campaign" is not worth tracking and excluding. Although I do believe companies should take the extra effort to avoid the annoyance and be a good citizen - particularly once they have identified "the one real person who actually bought something" among 99 other bots/crawlers/untraceable entities. Not all ad companies appropriately value the real person and the annoyance caused to them. Some do.
I'd argue the value was in it for media buyers. A lot of retargeted inventory is spent targeting people who are going to come back anyway. The agency then spins the numbers and presents it as if they were the only step in the process.
Some even take it a step further and conflate their retargeting with their prospecting numbers - there's a lot more uneducated advertisers out there for agencies to rip-off than educated ones.
Other than your standard ecommerce flows, online advertising is still an area of very murky attribution.
I am excluding the likes of companies like BlueKai (now Oracle) who facilitate trading the cookie to other companies for further targeting or completely different chain of ads based on a search click to another hotel chain
That said, all of this is getting killed by recent Safari and Chrome change of getting rid of all 3rd party cookies
Surrounded by content on advice "blue light"-- gets targeted for Blue-tooth lightbulb
So the only way it makes any sense is if Amazon is advertising on the chance that you’re going to cancel or return it in the future. Amazon certainly has those numbers. But do you think the chance you’re going to cancel or return it is so high Amazon should advertise the same product to you instead of something else you might buy?
It’s not either or - Amazon showed me grill accessories along with other models when I bought a grill. Regardless, I don’t really have a hard time believing that even a small chance of buying and canceling could make the expected value of showing that item higher than other items, even if that’s counter intuitive to you. I have a harder time believing that what’s shown is not well optimized and “obviously” wrong.
So yes, it is either/or.
Or, depending on the numbers, all grills or all accessories could also make sense. My broader point is that it’s perfectly possible the best strategy doesn’t line up with people’s intuition, which is why we use ML for recommendations in the first place.
But I really don't feel like this is a case where "intuition might be completely wrong" applies—buying a $400 grill, and then just going back and buying 3 more $3-500 grills, just isn't something that people do on a regular basis. Same for vacuum cleaners, refrigerators, etc. And yet Amazon does regularly advertise these things to us after such large purchases.
The only logical conclusion is that it's cheaper for them (in the short term, at least) not to update their ad/recommendation algorithm to take into account the nature of large one-time purchases.
> The only logical conclusion is that it's cheaper for them (in the short term, at least) not to update their ad/recommendation algorithm to take into account the nature of large one-time purchases.
Is that really the only logical conclusion? I’d think that “I don’t know people’s buying behavior on Amazon well enough to understand why this happens” should at least be a possible explanation.
ML is not some holy grail that always produces optimum algorithms. It depends heavily on inputs and on the assumptions, both explicit and implicit in the training data. My guess is Amazon does not have a "single-purchase item" model and it's not something the AI was trained to identify and so they aren't specifically excluded from groups of "similar-or-same" items.
I think it is far more likely that Amazon is optimizing the cost of development against the marginal benefit of not showing ads for single-purchase items than that this is somehow driving more sales.
1. Buy a large ticket item (e.g. furniture).
2. See an ad on Amazon for a different/nicer/better/faster shipped version.
3. Buy it.
4. Cancel the old order.
Is that a common pattern? No idea. But it’s not crazy to me that it happens often enough that it’s worth it to Amazon to show similar items to people who already bought them.They never advertised bags for said vacuum...
It makes sense to me that if I buy post hole diggers you should show me ads for work gloves or mailboxes or fences. Not more post hole diggers.
Weird example, I know.
As in if there are 10000 people that with browsing patterns similar to yours and non-trivial fraction of them one day show interest in a discount/service/trip then you are also likely to be interested.
It is the official explanation of how facebook/google can magically listen to conversation IRL that they were not supposed to hear: we simply are not as unique/unpredictable as we would like to think.
It was at least 5 years ago when I started to notice businesses "A/B testing themselves to death"
— "It makes no common sense, but increases sales,"
and then it does until it doesn't, and you are left with no idea what your client actually wants, and no idea how to get out of that "A/B tested to absurdity" situation
An A/B test isn't going to show that.
I don't want to fundamentally challenge the ŷ-centric approach approach of machine learning. But I do want to say that there's a lot of danger in taking it too far, and thinking that β̂ doesn't matter as long as ŷ is good enough. β̂ has a habit of not mattering until, very suddenly, it really matters quite a lot. And it's even more dangerous to think that an algorithm for generating formulas is somehow doing something intelligent just because you like what you're seeing in the ŷ department.
So, take the simplest possible model: You think that your dependent variable is linear in the dependent variable, and that the line passes through (0,0). That gives us a simplified version of the classic equation for a line, where we say that y is equal to x times some constant that indicates the slope of the line:
y = mx
In machine learning (or statistical inference), the variable conventions are slightly different, so it would look more like: y = βx
And then, since this is machine learning (or statistical inference, or whatever), we don't know the true equation, we're just working with an estimate we've inferred. We stick hats over the estimated values to make it clear that they're estimates: ŷ = β̂x
What I was describing, then, is which side of the formula you're focusing on. Traditionally, statistics has been all about scientific inference, and trying to figure out and characterize causal relationships. So you want to know "If I change x in some way, how will that affect the outcome, all else being equal?" β̂ is your tool for teasing those things apart. Machine learning is typically cast as being more about making predictions from observational data. ŷ is just the formal notation for describing those predictions.It's pretty common in machine learning to say, "We are working with observational data, we're not worried about causality, so we'll only pay attention to the predictions." There can be value to that approach, but only to the extent that you really don't care about causality. Which. . . surprisingly often, β̂ matters, even when the people who built the model were trying to ignore it. It becomes especially important when you're using the model to make decisions. For example, when a company is using a machine learning model to try and identify candidates for a position they're trying to fill, they should absolutely be paying attention to what the model thinks about people who liked Lane Bryant on Facebook, and how that compares to people who liked Jos A Bank, and whether they really want things like that to be playing into how they make hiring decisions.
Obviously they can help each other if you do things right and are lucky.
The ML just can’t figure her out. One day every ad is for dating sites, the next day it’s all babies and pregnancy, then engagement rings, wedding locations, more dating sites, a few days of houses to buy ... it’s like the algorithms can’t decide what life stage she’s in.
Dumb algos means freedom and privacy. Smart algos that actually work stuff out are dangerous and vile. Nobody gave them permission for that. Nobody complains about seeing ads not specifically tailored for them.
Or, it's working on figuring that out.
1. Generally not that advanced (yes I searched for hotels in Florida and so yes I’m probably taking a trip to Florida... don’t need fancy ML to figure that out.)
2. Not that smart. (Hey we actually got back from our Florida vacation two months ago... don’t need to keep seeing hotel offers for Florida!)
It totally makes sense that just “dumb” targeting (show people adverts related to the content on the page they are currently looking at, which requires no tracking) might actually work better than all the black magic products companies promote.