The article for this weeks discussion focuses on concentrated efforts to influence an items rating.
My proposed solution to deal with “trolls” influencing recommender systems, would be a system using popularity model vs average rating; if the normalized rating falls outside of 2 standard deviations (the item is too good or bad to be true), the item would fall into another dataset; the item would be given an internal neautral rating (2.5, .5, 5) and would then only be used via a content-based recSys. e.g. “Because you liked movies featuring ISIS and Santa clause, here is episode 17 of South Parks season 6”
The only example I can think of slightly related, would be the Dub the dew event hosted by mountain Dew, where users were asked to recommend a name for their new apple flavored mountain dew product line. Users with malintent made sure the no.1 recommendation would be “Hitler did nothing wrong”
Following link is NSFW
Link is NSFW : http://knowyourmeme.com/memes/events/dub-the-dew