IMO this is pretty solvable by looking at an account's purchase history too, but I don't think it's just as simple as blocking non-verified-purchase reviews.
Currently, companies pay people to buy items. They can keep those items, they just need to leave a good review. There are intermediaries who handle lots of sellers, so people buy a mixed bag of random garbage in exchange for the occasional review.
Yes, you can probably test for statistical anomalies, but I'm willing to bet that's quickly countered too - just have people buy occasional legit items so their profile is "statistically normal".
As far as I can tell, Amazon tries to fight that by keeping their ML model secret so scammers don't learn too quickly, but essentially, they're currently finding out what the Internet learned about SEOs manipulating search results.
Step 1 is getting a list of new products that are receiving a higher than expected number of 5-star reviews
Step 2, add it to a list of similar products.
Step 3, find accounts that all happen to purchase the same products in the same order.
Step 4, ban accounts and sellers.
What is higher than the expected number of 5-star reviews?
Don't those numbers vary widely depending on product category, brand, and newness of the product?
How does that work for things that fit into more than 1 product category?
That's left as an exercise for the reader ;)
But you could simplify this to new products, which if ungamed, would probably follow a predictable curve of discovery, as opposed to a games "instant 5 star" rating.
As Machine Learning is just a fancy way of doing statistics, you will totally not need it for this.
Worse still is that even taking a stance to not take a stance because you know it cannot be done will bring backlash as there are many who demand you take their stance, even allowing easy distribution of self chosen block and filter lists are not enough. There are many who demand the appliance experience.
Did they think I was being paid to write a negative comment, is that such a thing? or perhaps they don’t like negative comments either. I don’t know.
The same applies to links and Google. Buying/renting backlinks costs money, and still pretty much everybody that wants to rank in competitive fields does it.
They can for example provide a coupon that will reduce the price.
I don't think it'd do much good. There are already large groups that subsidize verified purchases to get 5 star reviews. I read an article about it, confirmed some investigation on Facebook of on my own.
Basically someone runs a Facebook group where sellers advertise free product, promising Paypal reimbursement of the purchase price in exchange for reviews. There are at least (or were, I haven't checked recently) hundreds of Facebook groups across many languages with thousands of members each doing these activities. The reviewers are randos who like free stuff, and I don't think anyone could detect them if they only casually participate in the review scams.
If you return a bad product, try and get a replacement to see if it was just bad luck, but also please do write a detailed review.
If you get a good product and it barely has reviews please review it.
That may well be me. I don't think that "meets requirements" should merit a 5. When I do reviews, I view a bare "meets requirements" as 3 or maybe 4, depending on the type of product. I want to leave some headroom to be able to point out products that really do excel.
Another ambiguity is whether the rating is on an absolute scale, or normalized for value (it's not a perfect product, but it's super-cheap).
I'm not sure this would help much. I've been asked by vendors to leave a review and if it's 5 stars, I will get something in return. I think for some people this would be a "well, nobody is getting hurt, right?" decision and they'd just do it.
I suspect that Amazon reviews are going to be even less useful now. Especially given things like this:
> Amazon does not provide many specifics about how a product’s overall star rating is calculated, other than stating that it is not a simple average but instead uses “machine-learned models” that take into account factors such as how recent the rating or review is and whether it was a verified purchase or not.