Information Cascades in Wikipedia Editor Discussions
Source: http://nitens.org/docs/qteso10.pdf
This article looks into the dynamics of Wikipedia editor discussions: it investigates “Article for Deletion” (AfD) discussions where editors discuss and vote on what to do with an article suggested for deletion (each editor can vote to Keep, Delete, Redirect or Merge the article.) From each AfD, we thus have access to a sequential list of editors’ votes. We can visualize each AfD then as being like a “worm” that turns left for each Keep vote, and right for a Delete vote. It’s interesting to see what kind of voting patterns emerge:
Articles with a “spiral” shape have near-unanimous voting records. These are potentially prime examples of information cascade: an information cascade would induce a voting pattern where each successive member is pressured to conform to the herd, producing a sequence of nearly identical votes.
Articles that follow a straight line, like the New Cold War, are controversial. It seems unlikely that information cascade is happening here since new ideas/positions continue to take sway throughout the voting period, as evidenced by the alternating Keep/Delete votes.
Articles that follow an “S” shape are particularly interesting since in these cases, we have a steady stream of Delete votes, but then somehow the trend breaks and is followed by a stream of Keep votes (or vice versa). Assuming there is cascade behavior, this would mean the users managed to break free of the cascade! While the paper didn’t go into depth on these cases, it would be interesting to further explore what linguistic or social features explain trend-breaking behavior (the features could be represented as additional private high/low signals in the cascade model.)
In order to gauge herd voting behavior, researchers tested the hypothesis that each vote was influenced by past votes in the sequence. They formalized the question as follows: split the list of votes into two parts, a and b. Does the expected number of Keep votes in part b depend on the number of Keep votes in part a? i.e. Are future votes independent of past votes, or not? Indeed, a t-test suggests that the answer is yes. This suggests that information cascades do occur in these discussions. However it does not account for the possibility that some articles really are more likely than others to deserve deletion, a.k.a. that all editors are really acting on their (correct) private information. Future work might explore this confounding. But the possibility that information cascades occur in AfD discussions has important social consequences: it suggests that more work should be done to allow users to make a decision independently of others, perhaps through a secret ballot. Furthermore, the presentation of arguments should not give the first word to loud, powerful users who have unfairly high sway at the beginning of a discussion.