Methods and applications for ontology-based recommender systems

摘要

Recommender systems are a specific type of information filtering systems used to identify a set of objects that are relevant to a user. Instead of a user actively searching for information, recommender systems provide advice to users about objects they might wish to examine. Content-based recommender systems deal with problems related to analyzing the content, making heterogeneous content interoperable, and retrieving relevant content for the user. This thesis explores ontology-based methods to reduce these problems and to evaluate the applicability of the methods in recommender systems. First, the content analysis is improved by developing an automatic annotation method that produces structured ontology-based annotations from text. Second, an event-based method is developed to enable interoperability of heterogeneous content representations. Third, methods for semantic content retrieval are developed to determine relevant objects for the user. The methods are implemented as part of recommender systems in two cultural heritage information systems: CULTURESAMPO and SMARTMUSEUM. The performance of the methods were evaluated through user studies. The results can be divided into five parts. First, the results show improvement in automatic content analysis compared to state of the art methods and achieve performance close to human annotators. Second, the results show that the event-based method developed is suitable for bridging heterogeneous content representations. Third, the retrieval methods show accurate performance compared to user opinions. Fourth, semantic distance measures are compared to study the best query expansion strategy. Finally, practical solutions are developed to enable user profiling and result clustering. The results show that ontology-based methods enable interoperability of heterogeneous knowledge representations and result in accurate recommendations. The deployment of the methods to practical recommender systems show applicability of the results in real life settings.