They wanted to print an appropriate coupon after someone paid for their shopping. "£1 off hamburgers next week" or similar, to get the person to return to the shop the following week.
They said a very important feature was to avoid giving a discount for a product someone has just bought or a closely related product. This would annoy customers who felt they'd missed out. Therefore, they were updating Neo4J with the just-made purchases and calculating the recommendation in the time it took to process the payment.
Maybe Amazon's chaotic mess of similar products makes this much more difficult than for a large supermarket, or maybe Amazon doesn't think the annoyance is a big deal.
I think amazon.com is no longer day one if it ever really was. I've seen at least one instance where they lowered price and refused to price match sale that hasn't even been delivered yet. I think they have data shows that not all people will go through the hassle to return and rebut and/or there's enough lock-in that amazon.com does not care about frustrating users.
Maybe the people who would be frustrated would go through the hassle to return and buy again. I mean if you have prime and you haven't opened the box, it is just two more days.
Yes, they're commonly bought together, but there's mostly, if not completely, a one-directional relationship-- someone buying 13 might want 14, 15, 16, but not the other way around.
This feels like it's just a matter of adding one more signed value to the "people who buy A buy B" record table... a weight indicating if the take is mostly A->B, B->A, or neutral.
a) they've got a model that has learned that people who buy X tend to buy X again, and
b) they don't have a model that can properly distinguish between classes of X for which this is true and classes of X for which it isn't. They are aware of the problem and have made several attempts to fix it over the years but the results have never been sufficiently compelling to incorporate into the actual recommendation model in prod