Trust in Online Collaborative IS - Community-Built Databases: Research and Development - page 350

Table 14.2 Mean and standard deviation of reputation values for admins,
good users, and blocked. Users for the three Reputation Models
Admins and good users Blocked users

Model 1 0.5416 ( 0.2407) 0.0926 ( 0.2091)

Model 2 0.7835 ( 0.1698) 0.1884 ( 0.2962)

Model 3 0.8180 ( 1514) 0.2216 ( 0.3128)

Fig. 14.2 Distribution of reputation for good users/admins vs. blocked users based on the three
models. The X -axis shows the reputation bins and the Y -axis shows the percentage of users in
the bins

outperforms the other two models at detecting blocked users, while Model 3 out-
performs the other two models at detecting good users.

14.3.1.2 Reputation and User Behavior

In this section, we consider the application of the three models to estimate the
reputation of all users by extending the previous analyses. We first estimate
reputation values for all the users of English Wikipedia. Figure 14.3 shows the
distribution of reputation values for the three models. Unlike Models 2 and 3, where
higher reputation users are more dominant, Model 1 yields a higher number of low
reputation users. This is a direct consequence of the prompt punishment of a user in
Model 1 after his/her contributed data are deleted. The decrease in reputation
punishment occurs in Model 1 regardless of the reason for the deletion or the
reputation of the deleter. Hence, it is very likely that Model 1 overly shifts good
users to the left. This is also confirmed by the results of the previous experiments
and the poor TPRs of Model 1, compared with Models 2 and 3.
In order to evaluate the predictive value of the proposed reputation models,
we run another experiment. In this experiment, we calculate the reputation of all
the users of English Wikipedia up to time t and analyze the users' behavior up to
time t . Then, in a second phase, we analyze their behavior after time t , and correlate
this behavior with the reputation values calculated before time t . Specifically, we
measure the statistical correlation between the reputation of the users at time t and
their behavioral indicators before and after time t . We process history revisions up