Computationally Modelling Trust: An Exploration

Trust has been extensively studied in fields as varied as economics [13,17], business and marketing [8,24], politics, e-commerce [1,14,41], psychology [7,42,50], sociology [27], medicine, nursing [15], and computing science [10,11,19,39,40,44]. Whilst many definitions of trust have been proposed, none has been agreed upon. However, there have been some good efforts in inventorizing the different definitions, and trying to distil what characterizes trust and what different dimensions there are to it [15,29]. As noted by [29], different researchers have defined different types of trust. Based on their review, [29] defined six types of trust. Trust in this paper is closest to the one they called Trusting Beliefs, defined in [29] as “the extent to which one believes (and feels confident in believing) that the other person is trustworthy in the situation”, with trustworthy defined as “willing and able to act in the other person’s best interests”. The computational model provided in this paper will make our view on trust more explicit. Trust is key to long-term relationships [22], and user trust has been shown to affect the success of a system, in terms of increasing sales, users’ likelihood to stay with and return to the system [8], use of information provided [32], and users’ willingness to pay more [1]. A computational model of trust is particularly useful when an adaptive system interacts with a community of users (like a group recommender system does). Simply optimizing trust is then not possible: actions aimed to increase the trust of some users may well decrease the trust of others. An accurate trust model would allow tailoring of system actions (such as selected appearance for an embodied agent, empathetic explanations, items) to maintain the trust of all users. In a social system, models of users’ trust can also be used when facilitating interactions between users. Many factors influence the trust of an agent x in another agent y. In this paper, we will consider the following: (1) direct experiences of x with y, (2) indirect experiences, namely the reported experiences of others (often referred to as reputation [40]), (3) intuitions of x about y based on stereotypes, (4) empathy between x and y, and (5) characteristics of x. Sections 2 and 4 discuss these factors, and incorporate them into a simple computational model of trust. Sections 3 and 5 report on two experiments that investigate some of the issues arising from the modelling and propose improvements. Section 6 concludes this paper.