Managing trust in e-environments

Trust is essential to effective collaboration, meaning that (business) partners choose each other and make decisions based on how much they trust one another. The way how to assess trust in e-environments is different from those in physical world, as there are limited indicators available in online environment. In e-environments, trust is evaluated with the usage of trust and reputation management systems (TRMS) that are based on different mathematical trust and reputation models. The underlying mathematical trust and reputation model should be reasonable in order to implement effective TRMS. The thesis is based on the Qualitative Assessment Dynamics (QAD) trust model, which considers human factors involved in trust reasoning and as such presents sound mathematical trust and reputation model. At first, we evaluated the QAD trust model based on simulations. We proposed different scenarios with agents that played service requester and service provider roles, where some of the service providers provide unsatisfying services. Further, we performed two set of experiments. The first set of experiments demonstrates the effect of introducing trust and reputation management system into environment and demonstrates the basic properties of trust evaluation defined within the QAD trust model. The QAD trust model defines different QAD operators that model an agent’s behavior and process of trust evaluation. The second set of experiments shows the properties and the effects of the QAD operators. We analyzed the behavior of the agents modeled with different QAD operators and explained the QAD operators properties based on comparison of the decisions they made. One of the prominent problems within trust and reputation management systems is the presence of unfair ratings, which has not been sufficiently addressed so far. To address the problem of unfair trust ratings we present the QADE trust model, which extends the QAD trust model with elements such as private trust vector, historical trust matrices, historical private trust vectors, agents’ general mindset, QADE operators, attacker agent and similarity function. As such, the QADE trust and reputation model assumes and models the existence of unfairly reported trust assessments. It considers human-centric nature of trust, where differently reported trust values do not necessarily mean false value propagation but can also imply differences in agents’ trust attitudes. Based on the QADE trust model, we provide the method to identify and filter out the presumably unfair trust values. The method is two-part. Firstly, a trust evaluator finds other agents in society that are similar to him. Similarity between agents takes into account pairwise similarity of trust values and similarity of agents’ general mindsets. Secondly, the trust evaluator black-marks reported trust values if they differ with theirs. Trust values that are black-marked by certain amount of agents are considered to be unfair and excluded from trust computation. We compared the effectiveness of methods to decrease the effect of unfair ratings based on simulations. We made the simulations in environments with varying number of attackers and targeted agents, as well as different kind of attackers – individuals and collaborative attackers. The results showed significant improvements of our proposed method with average improvements from 28% to 58% compared to the other most representative filtering methods by Teacy and Whitby. Trust ratings shared by agents in e-environments are subjective as trust evaluation depends on evaluator’s personal disposition to trust. As such, aggregation of shared trust ratings to compute an agent’s reputation may be questionable without proper consideration of rating subjectivity. So that all agents understand the ratings in the same way, the reported trust ratings should be adjusted to each agent individually. In the last part of the thesis, we address the problem of proper trust rating analysis and aggregation. We propose an extension of the QAD trust model, named QADES. We propose a novel Human-Oriented Method for Ratings Adaptation (HOMRA) that derives adjusted reputations compliant with the behavioral patterns of the evaluators. With HOMRA method, all participants have comparable opportunities to choose trustworthy agents in future transactions, regardless of their trust dispositions. The proposed HOMRA method finds the agents with similar trust dispositions, taking advantages of non-parametric statistical methods. After that, it computes the personalized reputation scores of other agents with the aggregation of trust values shared by agents with similar trust dispositions. The method derives the characteristics of agents’ trust dispositions implicitly from their past ratings and does not request them to disclose any part of their trust evaluation process, such as motivating criteria for trust assessments, underlying beliefs, or criteria preferences. We evaluated the performance of our method with extensive simulations with varying number of agents, different distributions of agents’ personality types, as well as simulations with different number of available trust ratings. The results showed significant improvements of our HOMRA method with average improvement of 50% over the Abdul-Rahman and 73% over the Hasan method, when comparing the efficiency of the methods depending on the number of agents in certain e-environment, the efficiency of the methods depending on the distributions of agents with different personality types and the efficiency of the methods depending on the availability of trust ratings