Application of generalized memory-based collaborative Recommender agents to service discovery

The introduction of the World Wide Web dramatically impacted our fundamental notion of information sharing, providing unparalleled awareness of both the power of information access and the penalty of information overload. Today's research on Semantic Web techniques focuses on the next step, a Service Oriented Architecture supporting automated sharing of services, both functionality and data. Personalized service recommendation tools utilizing user preference data would be extremely valuable in tailoring information access to the user. Much can be learned from the Recommender community about incorporating preference data into the retrieval process. However, it is critical that rigorous statistical techniques be maintained in combining results across sources that are not under the control of a single developer.; In this research we explore the impact of extending nonparametric techniques to preference-based collaborative ranking. In the process a unique nonparametric Filter-on-P-Weight-on-Rho algorithm is developed that avoids the assumptions of normality that are so common to traditional memory-based collaborative ranking algorithms. Using a standard Recommender data set we demonstrate a statistically significant improvement in the algorithm's accuracy over its parametric counterparts. In the process, we introduce several advances in Recommender evaluation techniques. Finally, a framework is established for inserting these more general Recommender techniques into a Service Oriented architecture to augment current content-based Discovery Services with a personalized collaborative discovery tool.