5 ms·

I do modeling for an advertising recommendation engine (not at Amazon). We do controls against both human-made rulesets and random, and we consistently outperfo

by cirgue 8y ago

I do modeling for an advertising recommendation engine (not at Amazon). We do controls against both human-made rulesets and random, and we consistently outperform both by a big margin. However, click through rate is always really low, and a model that massively increases revenue still serves irrelevant content most of the time, and still seems random from the POV of the user. The point of ML in advertising recommendation is to guess better in aggregate, not get it right all the time.

b_tterc_p 8y ago
What about human models plus baseline likeliness to click on any ad?
- manicdee 8y ago
  When the machines can predict high click-through rates, humans will be obsolete. What is even scarier is when ads don’t generate click-through rates at all, rather success will be measured by “subject A saw this propaganda” correlated to “subject A subsequently performed this action.”
  - b_tterc_p 8y ago
    Well basically yes. Cambridge analytica more or less measured how stupid you were and rationed out appropriately outlandish lies. What I’m asking is whether we can establish your baseline likelihood of clicking any ad, and then use human made models to determine placements. My hypothesis is that naive human models only lose out because they miss the low hanging fruit of avoid people who never click anythingf
    - manicdee 8y ago
      What if the reason click through are so low is that there is little demand created by marketing and only a relatively low rate of unexpressed demand? In my case, when I want something I will go find it. I spent three months looking for reviews of small, efficient cars before deciding on a VW Polo TDI. The day I went in to buy one, the salesman was only interested in selling stock they had in the floor that day (it was some kind of stocktake clearance) and none of those were TDI. He was extremely rude too. So I walked out of that dealership and bought a Mazda 2 Then I got three months of ads on various sites advertising Mazda 2. Never any ads for small, fuel efficient cars while that was what I was searching for on Google. The magic of targeted advertising completely failed to work. For the next few years, any ads for small efficient cars will be absolutely useless because I have no need for another one. And for some reason I am getting lots of ads for women’s clothing. I suspect this is based on assumptions about men in long term relationships looking for a bit on the side, but once again those ads miss their mark because I have never bought womens clothing, I do not visit sugar daddy dating sites, so I have no idea why they think I should be interested. It’s not just “people who don’t click anything,” it’s “people who aren’t interested in what you are selling,” and “the thing you are selling is not an impulse buy.” Nobody buys a second economy car as an impulse buy. A good deal on a 600mm Nikon F Mount might catch my interest though (but they’re all the same price as my car, so I don’t necessarily have that kind of money)

by cirgue 8y ago

b_tterc_p 8y ago

What about human models plus baseline likeliness to click on any ad?

manicdee 8y ago
When the machines can predict high click-through rates, humans will be obsolete. What is even scarier is when ads don’t generate click-through rates at all, rather success will be measured by “subject A saw this propaganda” correlated to “subject A subsequently performed this action.”
- b_tterc_p 8y ago
  Well basically yes. Cambridge analytica more or less measured how stupid you were and rationed out appropriately outlandish lies. What I’m asking is whether we can establish your baseline likelihood of clicking any ad, and then use human made models to determine placements. My hypothesis is that naive human models only lose out because they miss the low hanging fruit of avoid people who never click anythingf
  - manicdee 8y ago
    What if the reason click through are so low is that there is little demand created by marketing and only a relatively low rate of unexpressed demand? In my case, when I want something I will go find it. I spent three months looking for reviews of small, efficient cars before deciding on a VW Polo TDI. The day I went in to buy one, the salesman was only interested in selling stock they had in the floor that day (it was some kind of stocktake clearance) and none of those were TDI. He was extremely rude too. So I walked out of that dealership and bought a Mazda 2 Then I got three months of ads on various sites advertising Mazda 2. Never any ads for small, fuel efficient cars while that was what I was searching for on Google. The magic of targeted advertising completely failed to work. For the next few years, any ads for small efficient cars will be absolutely useless because I have no need for another one. And for some reason I am getting lots of ads for women’s clothing. I suspect this is based on assumptions about men in long term relationships looking for a bit on the side, but once again those ads miss their mark because I have never bought womens clothing, I do not visit sugar daddy dating sites, so I have no idea why they think I should be interested. It’s not just “people who don’t click anything,” it’s “people who aren’t interested in what you are selling,” and “the thing you are selling is not an impulse buy.” Nobody buys a second economy car as an impulse buy. A good deal on a 600mm Nikon F Mount might catch my interest though (but they’re all the same price as my car, so I don’t necessarily have that kind of money)