Shill reviews are a huge problem but I don’t see how an algorithm is going to find them. If 50 of my friends buy something from my company and then leave glowing reviews, how does Amazon know that we know each other?
Yes, I know I don’t have 5 friends, but work with me here.
If I were coding a shill-detection system for an online seller, I would begin with behavioral analysis of the user's actions before the purchase was even made. Did the user shop around, or just zero in on a specific product? Did they compare the product with others? Did they read reviews left by others? In other words, how different was the user's behavior from that of a typical buyer?
The system would also take note of how many of that item had been sold both historically and recently. A sudden spike in sales for an item that previously had a rather flat sales pattern, or of sales by a particular third-party seller, would raise suspicion.
Next the system would check whether the timing of the review seemed right. In most cases, vendors know exactly when a product was delivered, right down to the second. Reviews posted immediately upon receipt of the product would be considered more suspicious than reviews posted after a few hours or days, especially if the product was one that would require a bit of time to assemble, evaluate, etc., as determined by the average timing of reviews left for that product by other customers.
The content of the review would then be analyzed for excessive use of superlative adjectives and adverbs, either positive or negative. (Paid negative reviews are a real thing.)
The customer's review history would also be examined. Have all their reviews been either five-star or one-star reviews? That would be unusual enough to raise suspicion. Most people occasionally buy stuff that falls somewhere in between.
If enough suspicious factors were identified, a flag would be triggered for a more thorough human review than usual.
Note that none of the factors, by themselves or in aggregate, would result in a review being rejected or a reviewer being banned. All of the unusual behavior or patterns can be explained in legitimate ways. For example, a sudden spike in sales could be explained by advertising or by a positive post on an enthusiast forum for a product, and there are people who do know exactly what they want to buy when they first visit a site. So a human, not a robot, would have to make the final decision.
But algorithms could be used to narrow down reviews that smell fishy.
Rich