No abstract

Recommender systems that incorporate a social trust network among their users have the potential to make more personalized recommendations compared to traditional collaborative filtering systems, provided they succeed in utilizing the additional trust and distrust information to their advantage. We compare the performance of several well-known trust-enhanced techniques for recommending controversial reviews from Epinions.com, and provide the first experimental study of using distrust in the recommendation process.

Trust and distrust are two increasingly important metrics in social networks, reflecting users' attitudes and relationships towards each other. In this paper, we study the indirect derivation of these metrics' values for users that do not know each other, but are connected through the network. In particular, we study bilattice-based aggregation approaches and investigate how they can be improved by using ordered weighted averaging techniques, or through the incorporation of knowledge defects. Experiments on a real world data set from CouchSurfing.org demonstrate that the best operators from a theoretical perspective are not always the most suitable ones in practice, and that the sophisticated aggregation methods can outperform the more obvious bilattice-based approaches.

Recommendation technologies and trust metrics constitute the two pillars of trust-enhanced recommender systems. We discuss and illustrate the basic trust concepts such as trust and distrust modeling, propagation and aggregation. These concepts are needed to fully grasp the rationale behind the trust-enhanced recommender techniques that are discussed in the central part of the chapter, which focuses on the application of trust metrics and their operators in recommender systems. We explain the benefits of using trust in recommender algorithms and give an overview of state-of-the-art approaches for trust-enhanced recommender systems. Furthermore, we explain the details of three well-known trust-based systems and provide a comparative analysis of their performance. We conclude with a discussion of some recent developments and open challenges, such as visualizing trust relationships in a recommender system, alleviating the cold start problem in a trust network of a recommender system, studying the effect of involving distrust in the recommendation process, and investigating the potential of other types of social relationships.

Trust is not only essential in our daily lives, its importance in virtual networks should not be underestimated either. For example, we only use information found on the web when we trust the source, and we do not give our credit card details on sites we distrust. With the advent of the semantic web [2], intelligent web agents will be taking over more and more human tasks. Therefore, there is a growing need for computational models that can imitate the human notions of trust and distrust.Efficient trust models already play an important role in many intelligent web applications, e.g. question answering systems [9], P2P networks [7] and recommender systems [8]. Recent publications [5] also show an emerging interest in modeling the notion of distrust, but models that take into account both trust and distrust are still scarce [3,4,6].A trust network is a network in which the agents are connected by trust scores. Typically these networks are sparse. A fundamental problem in such networks is the determination of the scores of the agent pairs for which we did not receive an explicit score, i.e. propagation and aggregation of trust. The problem of trust propagation can informally be described as: if the trust value of agent a in agent b is p, and the trust value of b in agent c is q, what information can be derived about the trust value of a in c? Aggregation operators are needed to combine the trust values received from different trusted third parties (several propagation chains).Most of the existing approaches only take into account trust, and cannot distinguish between complete distrust (trust=0, distrust=1) and ignorance (trust=distrust=0). Besides, they deal with trust in a binary way: they assume that an agent is to be trusted or not and calculate the probability or belief that the agent can be trusted. But in reality, people often say 'I trust this person very much' or 'I rather distrust this person'. So there also is a need for representing partial trust.Therefore, we propose a model that represents both trust and distrust, as a matter of degree. To this aim, we draw (trust,distrust) couples from a bilattice-based square L 2 [1], where L is a complete bilattice. This representation allows to model partial validity as well as partial knowledge. In particular, this approach is powerful enough to model ignorance (no trust, but also no distrust) and inconsistency (trust and distrust at the same time).