A model of collaboration-based reputation for social recommender systems

Today's online world is one full of rich interactions between its users. In the early days of the web, activity was almost exclusively solitary, now however, users regularly collaborate with one another, often mediated by a piece of content or service. In offline communities, continued good behaviour and long-term relationship building leads naturally to good reputation, however online users often remain anonymous to their community and so trust building can be difficult to foster among community members. As such there has arisen a need for the system to calculate user reputation itself. Now, online reputation systems provide a variety of benefits to the platforms that employ them such as an incentive mechanism for good behaviour and improving the robustness of the platform. However, these systems are often based on ad-hoc activity metrics, and thus do not generalise to multiple platforms or different tasks.In this thesis we introduce a novel approach to capturing and harnessing online reputation. Our approach is to develop a computational model of reputation that is based on the various types of collaboration events that naturally occur in many different types of online social platforms. We describe how a graph-based representation of these collaboration events can be used to aggregate reputation at the user-level and we evaluate a variety of different aggregation strategies. Further, we show how the availability of this type of user reputation can be used to influence traditional recommender systems by combining relevance and reputation at recommendation time. A major part of our evaluation involves integration with the HeyStaks social search system, testing our approach on real-user data from the service.