Computing of trust in complex environments

In this paper, a novel information theoretical model for trust metrics in complex environments is proposed and analyzed. The contribution of this paper is two folds. In the first fold, syntax of actions is introduced and formalized in the context of disjunctive normal form over multiple Boolean variables, and random variables for describing an agent's behavior is formalized based on the notion of actions and auxiliary information (and auxiliary functions). The notion of trust is then formalized as a combination of a predicator and an evaluator of the current auxiliary information. In the second fold, we model contribution of a recommender's by a well studied notion mutual information of variables of behavior and variables of auxiliary information. We show that our trust metrics is sound in the sense that - the definition of trust metrics is transitive; if a node (form the point view of the subject) in a recommendation path is not trusted by the subject, the contribution of this recommendation path is zero; the longer the size of a recommendation path, the less trust value should be computed from individual recommenders along the path. The proposed trust evaluation system is useful in reputation systems, risk management, collaborative filtering and social networking services.