Completação fora de amostra em grafos de conhecimento

• Ano de defesa: 2023
• Autor(a) principal: Silva, Daniel Nascimento Ramos da
• Orientador(a): Porto, Fábio André Machado
• Banca de defesa: Porto, Fábio André Machado, Gomes, Antonio Tadeu Azevedo, Mattoso, Marta Lima de Queirós, Sampaio, Jonice de Oliveira, Valduriez, Patrick
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Laboratório Nacional de Computação Científica
• Programa de Pós-Graduação: Programa de Pós-Graduação em Modelagem Computacional
• Departamento: Coordenação de Pós-Graduação e Aperfeiçoamento (COPGA)
• País: Brasil
• Palavras-chave em Português: Teoria dos grafos; Grafos de conhecimento; Aprendizado por máquina; Redes neurais
• Área do conhecimento CNPq: CNPQ::CIENCIAS EXATAS E DA TERRA::MATEMATICA::ANALISE
• Link de acesso: https://tede.lncc.br/handle/tede/377

Knowledge graphs provide a semantic layer valuable for various applications. They repre- sent facts as a network of relationships between entities. However, many relevant facts may be missing from the graph, potentially impairing the performance of downstream applications. As a result, the development of knowledge graph completion strategies has proliferated in the last years, aiming to infer the truth value of relationships not observed in the graph. Overall, techniques based on representation learning have become the most common approach for completing knowledge graphs. However, many cannot perform inferences for emerging entities, not seen at training, which is incompatible with the evolving nature of knowledge graphs. In this scenario, a promising strategy, but not well investigated by researchers, uses the surrounding neighborhood of a fact — named query context — to infer its truth value. Given the above, we investigate the out-of-sample knowledge graph completion task, not limited to predictions for facts involving entities seen at training time. In detail, we develop and empirically evaluate a methodology based on query contexts and representation learning. First, we study the definition of this neighbor- hood and how it impacts the performance of learned models. Secondly, we develop neural network architectures for this task. Furthermore, we devise strategies for dealing with scalability problems inherent to this methodology. Finally, we carry out comprehensive experiments, which shed light on the challenges faced by the method and indicate that it is competitive with the state of the art.