Bayesian Inference in Trust Networks

Abstract

Trust has emerged as a major impediment to the success of electronic markets and communities where interaction with the strangers is the norm. Social Networks and Online Communities enable interaction with complete strangers, and open up new commercial, political, and social possibilities. But those promises are rarely achieved because it is difficult to trust the online contacts. A common approach to remedy this problem is to compute trust values for the new contacts from the existing trust values in the network. There are two main methods: aggregation and transitivity. Yet, neither method provides satisfactory results because trust networks are sparse and transitivity may not hold. This article develops a Bayesian formulation of the problem, where trust is defined as a conditional probability, and a Bayesian Network analysis is employed to compute the unknown trust values in terms of the known trust values. The algorithms used to propagate conditional probabilities through the network are theoretically sound and based on a long-standing literature on probability propagation in Bayesian networks. Moreover, the context information that is typically ignored in trust literature is included here as a major factor in computing new trust values. These changes have led to significant improvements over existing approaches in the accuracy of computed trust, and with some modifications to the algorithm, in its reach. Real data acquired from Advogato network is used to do extensive testing, and the results confirm the practical value of a theoretically sound Bayesian approach.