Reputation-Based Ranking Systems and Their Resistance to Bribery

Abstract

We study bribery resistance properties in two classes of reputation-based ranking systems, where the rankings are computed by weighting the rates given by users with their reputations. In the first class, the rankings are the result of the aggregation of all the ratings, and all users are provided with the same ranking for each item. In the second class, there is a first step that clusters users by their rating pattern similarities, and then the rankings are computed cluster-wise. Hence, for each item, there is a different ranking for distinct clusters. We study the setting where the seller of each item can bribe users to rate the item, if they did not rate it before, or to increase their previous rating on the item. We model bribing strategies under these ranking scenarios and explore under which conditions it is profitable to bribe a user, presenting, in several cases, the optimal bribing strategies. By computing dedicated rankings to each cluster, we show that bribing, in general, is not as profitable as in the simpler without clustering. Finally, we illustrate our results with experiments using real data.