Tournament-Based Reputation Models for Aggregating Relative Preferences

Online social networks often rely on reputation values to signal trustworthiness of a participant in the network. Reputation models have traditionally focused on scenarios where raters provide an absolute value rating to those they interact with. This paper investigates an alternate scenario, where a rater provides relative preference information among multiple alternatives, indicating which participant it prefers from a given set of participants. To model such scenarios, we propose an alternative to the traditional referral graph model, in the form of tournament models. We investigate two tournament-based trust models: the power ranking model and the fair bets model. We find that tournament models outperform standard referral network models, in predicting which participant will prove to be most preferred in future transactions. We also find that, in online communities where self-selection norms are not strongly enforced, so that a participant may participate in a transaction even when it knows it is unlikely to provide a satisfactory result, the fair bets model outperforms other models by a wide margin in identifying trustworthy participants.