GRAFT: A Distributed Recommendation Framework

Since the earliest human communities, reputation has been used by people to decide whether they should trust and interact with someone else. Traditionally, reputation was established through a person’s standing, word of mouth and their associations. However, with the increasingly widespread use of the Internet, this situation has changed. In particular, all of the normal cues that help to build reputation are missing. Even the concept of identity is blurred by the common usage of pseudonyms. In answer to this problem, many websites on the Internet have developed reputation systems that allow members to leave feedback about the performance of others in the execution of their duties. This accumulation of feedback about any individual can be used to characterise and predict their future behaviour in that context, allowing others to decide if they want to interact with that individual. Unfortunately, the information in each instance is limited to the narrow context of the website in which it was generated. Not only is the reputation information constrained in context, it also limits the potential scope of what can be determined about an individual. The information that could be collected about entities includes social, demographic and reputation-based information. These are collectively called recommendation information in this thesis. Collecting this recommendation information from multiple sources and contexts should provide a wider view by which an entity can be evaluated than reputation alone could produce. The combination of these multiple sources of recommendation information can be naturally extended in the development of novel applications in areas such as access control and web service composition. The GRAFT framework developed in this thesis encapsulates a paradigm shift in the way that reputation information is handled. It directly supports the collection and distribution goals by building a global distributed recommendation system that can be used to collect and make available recommendation information about both people and electronic services. This system can be used as both a drop-in replacement for existing systems, or it can be used to drive the consumption of recommendation information in novel new systems. Recommendation information can be collected from both traditional reputation sources such as Amazon and eBay, and non-traditional reputation sources such as social networks, providing flexibility in what can be collected and subsequently utilised by consumers. The derivation of reputation information from non-reputation sources including demographic and social information, and the subsequent ability to use this recommendation information in the description and evaluation of policies is unique to GRAFT. The major contributions of this thesis in the areas of reputation and reputation systems include the development of a reputation terminology, generalised models of reputation and reputation context, an extensive survey and taxonomy of reputation systems and a classification of existing reputation systems based on the taxonomy. This thesis also contributes an architecture for GRAFT, a prototype implementation of GRAFT showing its usefulness, and an evaluation that includes the results of a large number of simulation experiments showing how the architecture scales and handles both malicious peers and churn.