Semantic Recommender System

• Autor(a) principal: Kinaan, William
• Data de Publicação: 2016
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.22/11313
• topic: Semantic Recommender; Recommendation Systems; The Semantic Web; Ontology; Sistemas de Informação e Conhecimento

Though Content-based recommender systems proved to have better quality than Collaborative Filtering recommenders, the later is more used because the former suffers from complex mathematical calculations and inadequate data modeling techniques. Using Ontology(ies) to model the data allows machines to better understand both items and users’ preferences and thus not just suggesting better recommendations, but also providing accurate justifications. In this work we present a Semantic Recommender system that uses a novel way of generating recommendations depending on a Recommender Ontology that provides controlled vocabularies in the context of recommendations, and that is built upon the idea that not all classes and properties are important from item-similarities point of view. If the domain Ontology is annotated with the Recommender Ontology, the Semantic Recommender should be able to generate recommendations. As a result, the proposed system works with any domain data. Thanks to The Semantic Web standards. The proposed mathematical model takes into consideration, in addition to items’ features and users’ profiles, the context of the users and the temporal context, so some items, as an event’s ticket, should never be recommended if the event is over, and should get more presence before the event. The Recommender Ontology grants business owners a way to boost the recommended items according to their needs. This guarantees more diversity, which satisfies the business requirements. For the experiments, we have tested the proposed solution with many domains including movies, books, music, and with a real business company. We got 55% accuracy when testing on a movie domain though we knew just one feature about the movies. The main limitation we have faced is the absent of a content-based domain case that contains ABox, TBox, and ratings together.