Examinando por Autor "Gonzalez-Fierro, Miguel"

Mostrando 1 - 1 de 1
  • Miniatura
    Publicación
    Automatic Generation of Domain Knowledge Graph for Recommendation Systems using Open Source Resources
    (Universidad Nacional de Educación a Distancia (España). Escuela Técnica Superior de Ingeniería Informática. Departamento de Lenguajes y Sistemas Informáticos, 2020-09-01) Sanz Olive, Almudena; Albornoz Cuadrado, Jorge Carrillo de; Gonzalez-Fierro, Miguel
    A group of state-of-the-art recommendation algorithms using Knowledge Graphs are RippleNet [1] or KGAT [2]. However, the main bottleneck to benchmark and work with these algorithms is that they used Microsoft Satori 1 or Freebase 2 (now Google Knowledge Graph 3) as their KG. As Satori and Google Knowledge Graph are commercial KGs and not open source, it's not possible to replicate the results found in the papers on new data, or use these solutions in real applications without paying for access. This limitation is critical in the development of this area of RS. There are researchers working on algorithms and applications, whose results cannot be replicated openly by the science community, hence, going against the scientic method. As researchers, one of our main goals is to make science accessible and replicable for all the scientic community, empowering the development of new knowledge areas. To ll this gap, this Thesis presents a system to generate domain adapted Knowledge Graphs using open source information from Wikidata. These Knowledge Graphs can be used in hybrid Recommendation Systems, that use linked knowledge on the items as side information, combining both Collaborative Filtering and Content Base Filtering strategies. The results show that the proposed system is able to create domain adapted Knowledge Graphs from open source information for recommendation datasets, and that the KGs generated are able to compete with their commercial versions. We have shown that our proposed system creates smaller KGs that are more domain adapted, that have a similar eciency in downstream tasks of Recommendation Systems than commercial KGs. This could be specially relevant in systems that require faster computational times that can be achieved with smaller KGs, as real-time systems, or to save computational cost in high-scale systems. The system ca be used as a reference to evaluate the state-of-the-art algorithms in future works in the area.