Interesting NYTimes’s article (if you don’t want to register, use BugMeNot where you can find shared login and password pairs). Mikhail Gronas discovers that “reviewers gave more five-star reviews than two-star reviews, creating an upward sloping curve”. (…) “But the most telling variable is the one star rating. Professor Gronas found that books high on what he called the “controversiality index” are given almost as many one-star as five-star ratings, creating a horseshoe-shaped curve. As it turns out, these books also tend to have high sales.”
I’ve found these patterns analyzing Epinions.com ratings and trust statements (chech the graphs’ on the paper (pdf)) but actually I don’t think they are that surprising: they seem pretty obvious and I just reported them passing by.
What is really depressing is that Dartmouth is now in the process of patenting software that will be used to determine the “controversiality index”.
I’m happy that in Europe we are still fighting against a so-stupid-policy of being able to patent everything, no matter how trivial it is. In this case the controversiality level of a book is something like “if a book received as much 5 ratings as 1 and if the 5 and 1 ratings together are the vast majority of ratings and if the number of received ratings is over a threshold (probably depending on release time), then the book is controversial” (putting it in formula that produces a controversiality value would require 10 minutes at most).
By the way, I’m currently working on the concept of controversiality of users and hopefully a paper is on the way. Controversial users are users who are trusted by many and distrusted by many. (Bush is a good example, but this can happen to highly visible persons in general). The idea is that Local Trust Metrics make sense expecially for highly controversial users (for example, users who are trusted by more than 200 users and DIStrusted by more than 200 users in the community). For those users, it does not make sense to predict a trust value of 0.5 saying that you should trust this user as 0.5 but, instead, to predict you should trust this controversial user as 1 if, for example, all your friends trust her and 0 if all your friends distrust her.