> Machine learning algorithms are getting sophisticated enough to figure out whether a batch of 100 reviews is mostly talking about can openers or garlic.
This barely needs ML. Amazon knows, a priori, what product the reviews were for. The fact that garlic reviews can be repurposed for can openers, regardless of their actual text, is the problem.
The real underlying problem is wrong incentives. Amazon makes money when you buy that crappy can opener.
how do you clarify a valid product update? Aeropress changed the plastic used at some point in ~2013 but didn't publicize it widely. Should the Amazon listing start over from scratch for that type of change?
Apple reviews start over unless you explicitly ask for all versions. This is mildly obnoxious but works okay. Amazon could easily show reviews for the current product and a nice display of predecessor products and their reviews.
This would only work with the assumption that the owners of the product listing would update the version of the product in good faith. It would only be in their interest to do so if the current version of the project would get better reviews than the older version, and what I would imagine normally happens is the old version is made better to get better reviews and is then modified to be cheaper and worse later on.
Huh? If someone acquires a garlic listing and uses it to sell can openers without updating the listing, it won’t work, because everyone placing orders will think they’re buying garlic, not can openers. The scam only works because the listing matches (approximately, anyway) the sold product. If you start sending can openers labeled as garlic to customers, your return rate will be very high and there won’t be any profit.
This barely needs ML. Amazon knows, a priori, what product the reviews were for. The fact that garlic reviews can be repurposed for can openers, regardless of their actual text, is the problem.
The real underlying problem is wrong incentives. Amazon makes money when you buy that crappy can opener.