Trust in Online Collaborative IS - Community-Built Databases: Research and Development - page 355

revisions. Because reverts happen very frequently in Wikipedia, ignoring the side
effect of reverts can result in significant numbers of incorrect assignments of token
ownership.

14.3.2.2 Stability of Edits

For the purpose of this study, user reputation is estimated by looking at the stability
of the content he/she contributes. To estimate the stability of the content, we track
the tokens inserted by a user up to the last revision of the page to see how many of
these tokens are deleted. In some of the related work in the literature, the tracking
process has been more limited, for instance, by tracking inserted tokens only up to a
limited number of successive revisions and therefore missing some deleted tokens.
For example, the authors in [ 37 ] use up to the tenth successive revisions. Our study
of Wikipedia shows that 37% of the deletes happen after the tenth revision. Hence,
ignoring this fraction of deletes may lead to reputation estimates that are less
accurate. For the purpose of this study, user reputation is estimated by considering
the stability of inserts only. One may argue that although the number of deletes is
considerably smaller than the number of inserts, there is some information in the
stability of the deletes too, and one ought to be able to use this additional informa-
tion to derive even more accurate models of reputation. To see if the stability of
deletes can improve the accuracy of the models, we reformulate our simplest model
(Model 1) by considering the stability of deletes. We define Model 1 0 as follows,

n i n d ð

ð t Þþ t Þ

R i ð t Þ¼

Þ ;

(14.4)

N i N d ð ð t Þþ t

where n d ( t ) is the number of good quality deleted tokens and N d ð

is the total number
of deleted tokens after time t . We tested Model 1 0 as a classifier on admins and vandals
and the results showed that Model 1 0 has lower AUC (0.84) thanModel 1. Interestingly,
this observation is consistent with the result of another study [ 5 ], which shows that
delete and proofread edits have little impact on the perception of top contributors in
Wikipedia. In other words, there does not seem to exist any significant correlation
between an author's reputation and an author's number of deletes in the wiki pages,
but, in contrast, there is a very strong correlation between an author's reputation and an
author's number of insertions.

t Þ

14.3.2.3 Dynamic/Nondynamic and Individualized/Nonindividualized
Reputation Measures

One of the advantages of the models presented here is that they assign individualized
and dynamic reputation values to both anonymous and registered users. This is not
the case in some of the related work published in the literature. For example, the
authors in [ 42 ], use nondynamic and nonindividualized reputation values for the