MetaLens: A framework for multi-source recommendations.

摘要

In a world where the number of choices can be overwhelming, recommender systems help users find and evaluate items of interest. They do so by connecting users with information regarding the content of recommended items or the opinions of other individuals. In this thesis, we focus on a new class of recommender systems called meta-recommenders. Meta-recommender systems build on existing recommender technologies by giving users control over the combination of rich recommendation data to yield more personalized recommendations.; The work presented in this thesis makes several significant contributions to the field of recommender systems. We begin by considering the technologies used in creating recommender systems and the variety of ways these technologies are applied and recommendations presented in e-commerce recommender applications. We use this information to create a taxonomy for recommender applications in e-commerce. We also consider correlations between the recommender application models used to recommend products and the sites that choose to implement them.; Next, we introduce meta-recommenders and present the MetaLens Recommendation Framework. This framework serves as a model for how meta-recommenders collect data and generate recommendations that users find understandable, usable, and helpful. A series of controlled use experiments indicate that users want these systems to provide recommendation data alongside the recommendation. Furthermore, when appropriate, users want control over which data is displayed.; Implementation studies show the development of three different recommender systems built within this framework. Analysis of public use of these systems demonstrates that users like, and often prefer, these systems to more “traditional” recommenders. While acceptance comes at a slow pace, users who customized a system were more likely to return to use the system again. Finally, while the quantity and type of recommendation data preferred varies widely from user to user, analysis demonstrates that users want access to as much recommendation data as possible. All told, these results provide a meaningful foundation for the design of future meta-recommenders.