Difference between revisions of "Peer-to-Peer Accountability Enforcement/mechanism/CW"
| Line 1: | Line 1: | ||
'''Credibility Weighting''' (CW) means that all ratings do not figure equally in the aggregate. A rating from a user with a high credibility rating will count more than a user with a low credibility rating. | '''Credibility Weighting''' (CW) means that all ratings do not figure equally in the aggregate. A rating from a user with a high credibility rating will count more than a user with a low credibility rating. | ||
| + | ==Pitfalls== | ||
| + | Although it would seem logically consistent, it's arguable that users with ''negative'' credibility ratings must not have ''negative'' influence on the aggregate, as this would give them just as much power to "game the system" by voting ''against'' accounts they actually want to boost. Weighting should therefore either bottom out at zero, or else should follow some curve that never quite goes negative (e.g. log() of the user's rating). The ''magnitude of influence'' must be some positive function of the user's credibility. | ||
| − | ' | + | PCV may negate this effect, however, since any given user can't predict who will rate them negatively (and it may be a good idea to deny users information, either globally or selectively, about the ratings they have received from others). Further analysis is needed. |
| + | |||
| + | Note that it might be confusing if giving someone a negative rating gave them even a ''small'' amount of positive influence, so we want to avoid that as well. | ||
Latest revision as of 11:19, 24 October 2017
Credibility Weighting (CW) means that all ratings do not figure equally in the aggregate. A rating from a user with a high credibility rating will count more than a user with a low credibility rating.
Pitfalls
Although it would seem logically consistent, it's arguable that users with negative credibility ratings must not have negative influence on the aggregate, as this would give them just as much power to "game the system" by voting against accounts they actually want to boost. Weighting should therefore either bottom out at zero, or else should follow some curve that never quite goes negative (e.g. log() of the user's rating). The magnitude of influence must be some positive function of the user's credibility.
PCV may negate this effect, however, since any given user can't predict who will rate them negatively (and it may be a good idea to deny users information, either globally or selectively, about the ratings they have received from others). Further analysis is needed.
Note that it might be confusing if giving someone a negative rating gave them even a small amount of positive influence, so we want to avoid that as well.