The State of the Art in Trust and Reputation Systems: A Framework for Comparison
Zeinab Noorian, Mihaela Ulieru
2010
Journal of Theoretical and Applied Electronic Commerce Research
We introduce a multidimensional framework for classifying and comparing trust and reputation (T&R) systems. The framework dimensions encompass both hard and soft features of such systems including different witness location approaches, various reputation calculation engines, variety of information sources and rating systems which are categorised as hard features, and also basic reputation measurement parameters, context diversity checking, reliability and honesty assessment and adaptability
more »
... h are referred to as soft features. Specifically, the framework dimensions answer questions related to major characteristics of T&R systems including those parameters from the real world that should be imitated in a virtual environment. The proposed framework can serve as a basis to understand the current state of the art in the area of computational trust and reputation and also help in designing suitable control mechanisms for online communities. In addition, we have provided a critical analysis of some of the existing techniques in the literature compared within the context of the proposed framework dimensions. Given that T&R systems are context sensitive, the design of different existing models and systems has been dependent on the target domain and the related specific requirements. In the following, we review some of the available systems and discuss how they are able to fulfill their goals. We attempt to select the T&R systems with different approaches and techniques in dealing with the intrinsic challenges of the open environment. More explicitly, the chosen T&R systems have distinguished features in dealing with inherent dynamicity of the open environment, evaluating the honesty and reliability of participants, and calculating the reputation score. Such diversity enables readers to obtain decent understanding about existing literatures in trust and reputation systems and observe their applicability in virtual community. FIRE Model In the FIRE model [6]-[7], trust is evaluated within the context of a different number of information components: 1) Interaction Trust (IT) that is built from the direct self experience of an agent with the other agents; 2) Witness Reputation (WR) that is based on the direct observation of an agent's behavior by some third-party agent; 3) Certified Reputation (CR), being one of the novelties in the FIRE model, consists of certified references disclosed by third-party agents. Such information is made available upon request of an inquiring agent. The CR component is desirable in the absence of direct interaction and when witnesses are self-interested and reluctant to share their experiences. Moreover, the use of CR enables agents to be freed from the cost of locating witnesses while their confidence rate of the anticipated trust value in not compromised. 4), the last component is Role-based Trust (RT), which models the trust across predefined role-based relationships between two agents, e.g., (owned by the same Journal of Theoretical and Applied Electronic Commerce Research company, friendship relationship, team-mate relationship) [7] . In this case, by defining and updating these roles in open Multi-Agent Systems (MAS) as well as assigning the expected trust value and belief strength (of relying agent) on them, RT is able to contribute in trustworthiness prediction for future interactions. It is worthwhile to mention that the significance of each component in the composite trust formula is adjusted automatically according to unforeseen changes in the environment. In this trust model, each component owns a trust formula with relevant rating weight function to determine the quality of ratings tailored to its responsibility. For instance, it seems sufficient for IT to design the weight function solely based on the recency of ratings whereas WR and CR have to take the credibility of rating into account as well. To address this requisite, FIRE has developed a mechanism to filter out the inaccurate reports revealed by unfaithful witnesses and penalises them accordingly. In so doing, it defines an inaccuracy tolerance threshold (L) to specify the maximal permitted differences between the actual performance and witness rating. Credibility of each rating is tuned to be inversely proportional to the differences, i.e., the higher the differences are, the lower the credibility [5] . Furthermore, the FIRE model defines a reliability measure to calculate the confidence level of an agent in believing that another agent can perform as expected. In general, it provides two types of reliability: rating reliability, which depends on the number of available ratings with high values, which depict the expected performance of the target agent. The other type is deviation reliability, which intends to examine the volatility of the target agent in accomplishing an agreement. Basically, it calculates the deviation of ratings around the produced expected value [5] . Intuitively, if the target agent showed an inconsistent behavior while countering a different requesting agent, its reliability value will be gradually affected negatively. Note that the FIRE model inherits IT and multiple-criterion rating systems from the REGRET [26] reputation system and for the purpose of seeking and locating the relevant witnesses in WR, it is inspired by the decentralized approach of Singh and Yu's referral network [34] and implements a variant of their system. to estimate the total belief regarding the trustworthiness of a particular agent, the requesting agent combines a local belief in conjunction with third-party testimonies to achieve a more accurate evaluation. untruthful agent could considerably affect the reputation of the queried SP by providing a huge number of unfair ratings. This problem arises because of its method of reputation combination, which is based on a simple summation of all the provided opinions. To rectify this, TRAVOS adopts techniques to reduce the amount of ratings unless the accuracy degree of the opinion provider is very high. reputation aggregation method. It is noteworthy to mention that REGRET examines the truthfulness of information in general and does not differentiate between dishonest third-parties and incompetent but honest ones (R7b), (R7d). Moreover, it does not describe any method to update the weight parameter of witnesses after actual interaction results have been obtained. The other significant characteristic of this T&R system is its ability to support multiple-criterion rating system (R2a). As already mentioned in Section 2, REGRET is aspect-oriented and records reputations linked to a single behavioural aspects of a contract [24] . For instance, in case the contract consists of multiple criteria, it specifically inquires regarding particular criterion rather than a general reputation value. Alternatively, in order to calculate overall ratings, REGRET enables each participant to design an ontological structure of the contract suited to its requirement and weights each aspect proportionally. This feature addresses the criteria similarity rate (R5a) in a manner which leads to more precise prediction of the reputation value.
doi:10.4067/s0718-18762010000200007
fatcat:gpnldavetrdc3jaclji2mycl4e