Análise da técnica deep forest para o problema de aprendizado de ranqueamento

• Ano de defesa: 2022
• Autor(a) principal: Rocha, Lucas Elias Cardoso
• Orientador(a): Rosa, Thierson Couto
• Banca de defesa: Rosa, Thierson Couto, Sousa, Daniel Xavier, Rocha, Leonardo Chaves Dutra da, Canuto, Sérgio Daniel Carvalho
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Goiás
• Programa de Pós-Graduação: Programa de Pós-graduação em Ciência da Computação (INF)
• Departamento: Instituto de Informática - INF (RG)
• País: Brasil
• Palavras-chave em Português: Recuperação de Informação; Learning to rank; Métodos ensemble
• Palavras-chave em Inglês: Deep forest; Information retrieval; Learning to rank; Ensemble
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://repositorio.bc.ufg.br/tede/handle/tede/12087

Learning to Rank (LeToR) is a specialization of the ranking problem within the Information Retrieval (IR) field of study. In LeToR, machine learning algorithms are used to produce an ordered list of objects. The relative position of objects is given according to their degree of relevance or importance, depending on the problem application. A strategy present in state-of-the-art algorithms to handle LeToR is tree-based \textit{ensemble} methods. LambdaMART exemplifies this strategy and shows good results compared to other models, including deep neural networks, placing itself as state-of-the-art in LeToR. A tree-based \textit{ensemble} method not yet studied in LeToR, the Deep Forest, aims to perform deep learning without use deep neural networks. To do so, Deep Forest applies a layer-by-layer processing and an attribute transformation within the model through a \textit{ensemble} of \textit{Random Forest}. Due to the good performance shown by Deep Forest in several tasks and observing the good applicability of tree-based \textit{ensemble} methods in Learning to Rank, it is coherent to study the application of Deep Forest under the LeToR perspective. Having this general objective, the present work experimentally investigates Deep Forest aspects such as hyperparametrization, possible improvements of the original model, the model's behavior by bias and variance and the comparison with deep neural networks, all in the context of LeToR. It is expected that this investigation offers, in addition to the results of the application of Deep Forest in LeToR, an analytical view of the behavior of \textit{ensemble} models in LeToR and a comparative analysis of the results with deep neural networks.