Can you elaborate a little bit more on this, what do you optimize, what statistics do you use, etc
> We also use statistical models for ad optimization
Essentially, the problem (like all other ad opt. problems) boils down to estimating the expected value of a particular click. This is a particularly challenging/weird problem in this space because unlike other segments of ecommerce, you can't successfully optimize for correlated metrics such as engagement. So, it's a heavily imbalanced problem (i.e. few positive examples, lots of negative examples). In addition to that, the buying characteristics of specific products are heavily related to one another so sales on one affect how we advertise the other. There are a number of other subtleties discovered over time.
Our software produces good bid estimates despite these characteristics.
Essentially, the problem (like all other ad opt. problems) boils down to estimating the expected value of a particular click. This is a particularly challenging/weird problem in this space because unlike other segments of ecommerce, you can't successfully optimize for correlated metrics such as engagement. So, it's a heavily imbalanced problem (i.e. few positive examples, lots of negative examples). In addition to that, the buying characteristics of specific products are heavily related to one another so sales on one affect how we advertise the other. There are a number of other subtleties discovered over time.
Our software produces good bid estimates despite these characteristics.