https://douglas-fraser.com/FakeReviews/index.html https://douglas-fraser.com/FakeReviews/index.html I did my dissertation on this topic - text only analysis. I

https://douglas-fraser.com/FakeReviews/index.html https://douglas-fraser.com/FakeReviews/index.html I did my dissertation on this topic - text only analysis. I used a dataset that was commonly used in the beginning, but there are some issues with it. I plan to extend this to real reviews, as in 80 million Amazon ones (when I get the time). Text based features are useful, but non-text based ones are even more so. Even spamming groups can be detected; at least there has been research into that. Combining all the techniques in an ensemble would be productive - but is it really in Amazon's interest? My sense is whatever they do, they pick the low hanging fruit and trying to process every review that comes in would require a lot of CPUs perhaps. But stuff like floods of reviews for new products that are fairly similar should be easy to detect. Perhaps they are relying on Fakespot and reviewmeta to do the heavy lifting.