jbrnbrg
June 14, 2018
The commercial recommend system I've selected is one from Amazon's that has generally lead me to purchases I've been happy with: the Customers who viewed this item also viewed recommendations. Below, I've included an example of recs I received on a mechanical pencil I purchased last month:
The way this rec is made is given in the description - customers who looked at this same item also checked out these items - this is an example of Item-Item Filtering. It's not uncommon to see products that are category adjacent in these selections - e.g. for pencils they may show erasers and notebook planners - "pencil-adjacent" type products.
I suspect these recs are generated by querying user-sessions that included the given item among items reviewed during the session. It is possible that this is based on the idea that shoppers may be reviewing related/comparable items based on similar features or price. Sometimes, I'll be served some recs containing products that I am unfamiliar with but it's infrequent althought I suspect this is a drawback of this method.
This article discusses coordination between many users to flood a rec. system with 1-star reviews with the intention of damaging sales/reputations of the target.
There are a couple of options that spring to mind -- some of them potentially business-ending depending on the site -- for attacks of the nature described in the article:
Item 1 seems like it could be implemented with little disruption to existing-user experience but could be damaging to new-user's experience.
My idea for item 2 came from MetaFilter.com - "a community web log." Although it's not a rating site, I believe their new-user requirements would be effective in this case. They require a $5 life-time membership fee for the privilege of posting and they also require the user make a few minor posts before they can start sharing links. While it increases the quality of the content on the site, I recognize that adding a fee of any amount will almost certainly reduce new-user traffic.
While the author makes reference to some examples that indicate that these reviews have little impact on actual sales, I know from personal experience that a low rating will make me avoid a movie i.e. I generally avoid those with IMBD scores of 6 and under so there is certainly an impact.
While not directly related to a traditional "review" website, one recent example of a rec. system result's being altered through a crowd-sourced initiative or protest was when pro-Trump Reddit users figured out they could alter Google image results in the ramp-up to the 2016 election ( Link - warning: some may find some of this content disturbing - I know I did).
Another persistent and troubling trend I have noticed is the alteration of Google's auto-complete algorithm to produce politically motivated, vile, or outright false suggestions that lead to similar content. This article, again from election ramp-up, touches briefly on the issue after Google began to address it but a Feb.'18 article from Wired suggests it's still a problem (again, content may be disturbing to some.)
With the way folks in today's age find and consume information, addressing these issues is vitally important from a business perspective and also from the non-trivial perspective of maintaining a shared, cogent reality supported by factual information.