Evaluating Trust from Past Assessments with Imprecise Probabilities: Comparing Two Approaches

Abstract

In this paper, we consider a trust system where the trust in an agent is evaluated from past assessments made by other agents. We consider that trust is evaluated by values given on a finite scale. To model the agent trustworthiness, we propose to build imprecise probabilistic models from these assessments. More precisely, we propose to derive probability intervals (i.e., bounds on singletons) using two different approaches: Goodman’s multinomial confidence regions and the imprecise Dirichlet model (IDM). We then use these models for two purposes: (1) evaluating the chances that a future assessments will take particular values, and (2) computing an interval summarizing the agent trustworthiness, eventually fuzzyfying this interval by letting the confidence value vary over the unit interval. We also give some elements of comparison between the two approaches.