Recommender systems -- Interest graph computational methods for document networks.

摘要

Recommender Systems are now available in a number of online locations to help users find the reference information they need quicker and with greater accuracy. Document Networks are candidates for this technology to help researchers find research information which pertain to subjects in which they have an interest. Document networks are Bibliographic databases containing scientific publications, preprints, internal reports, as well as databases of datasets used in scientific endeavors such as the World Wide Web (WWW), Digital Libraries, or Scientific Databases (Medline). This Dissertation looks in detail at Document Networks and has chosen Semantic Medline for its case study. Semantic Medline supports thousands of medical researchers who wish to find available citations which pertain to a specific research interest from over 20 million medical research publications. I review Semantic Medline in some detail as well as Recommender Systems and how these systems are constructed and evaluated. So, the hypothesis is these new approaches will improve Document Network recommendations once implemented. The Dissertation first defines the requirements to improve Document Network recommendations. It then evaluates a host of algorithmic and technical approaches to the problem, selects the best candidate approaches, and a technical platform for evaluation is built to test these optional approaches using the actual Semantic Medline database loaded on a graph database engine. The original Semantic Medline is implemented with a more traditional database approach using MySQL queries to access and bring forward citations for search scenarios. This Dissertation uses new graph tools from social network technology to do the same thing and to evaluate these improved approaches to improve the recommendation accuracy and novelty. After a number of alternative approaches are tried, re-tested, and optimized, the best of the algorithms optimized for Document Networks are found and the original hypothesis is proven while also meeting the requirements. The results are interesting and can lead to greatly improved capabilities for Semantic Medline and for Document Networks in general.