Why not use a machine learning algo that trains itself based on removed reviews?

[Austin 0]

It seems like that kind of algorithm would be able to most accurately weight reviews across Amazon. Is there a reason you don't do this?

[ReviewMeta 35]

Hi Austin-

Thanks for chiming in.  We do use some forms of machine learning, but we're not just looking at the deleted reviews on Amazon.  One of the issues with strictly using machine learning based on deleted reviews as the answer key is that it's not a perfect answer key.  We can't say for certain that all undeleted reviews are honest and all deleted reviews are dishonest. 

Sometimes a user deletes their account and all their reviews.  Sometimes a company pressures a customer to remove their review.  Sometimes the reviews violate a policy on Amazon that we don't necessarily agree with - for example, you're not allowed to call out other fake reviews in your review.  So those reviews will often get deleted.  Technically they are against Amazon's TOS, but we don't consider those to be "fake".

Furthermore, we don't actually have the ability to collect every single deleted review.  We only know if a review is deleted when we collect it at one point, then check back on a later date and see it's no longer there.  So if a review is only live on Amazon for 1 day, it's unlikely that we'll ever see it.

We did a blog post on the topic of gauging our accuracy based on the deleted reviews a while back, might be worth checking out: https://reviewmeta.com/blog/how-accurate-is-reviewmeta-com/