Estudo comparativo de comitês de sub-redes neurais para o problema de aprender a ranquear

• Ano de defesa: 2023
• Autor(a) principal: Ribeiro, Diogo de Freitas
• Orientador(a): Rosa, Thierson Couto
• Banca de defesa: Rosa, Thierson Couto, Sousa, Daniel Xavier de, Canuto, Sérgio Daniel Carvalho, Martins, Wellington Santos
• 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 (RMG)
• País: Brasil
• Palavras-chave em Português: Recuperação de informação; Aprender a ranquear; Aprendizado por comitê; Comitê de sub-redes neurais
• Palavras-chave em Inglês: Information retrieval; Learning to rank; Ensemble learning; Sub-networks ensemble
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO
• Link de acesso: http://repositorio.bc.ufg.br/tede/handle/tede/13123

Learning to Rank (L2R) is a sub-area of Information Retrieval that aims to use machine learning to optimize the positioning of the most relevant documents in the answer ranking to a specific query. Until recently, the LambdaMART method, which corresponds to an ensemble of regression trees, was considered state-of-the-art in L2R. However, the introduction of AllRank, a deep learning method that incorporates self-attention mechanisms, has overtaken LambdaMART as the most effective approach for L2R tasks. This study, at issued, explored the effectiveness and efficiency of sub-networks ensemble as a complementary method to an already excellent idea, which is the self-attention used in AllRank, thus establishing a new level of innovation and effectiveness in the field of ranking. Different methods for forming sub-networks ensemble, such as MultiSample Dropout, Multi-Sample Dropout (Training and Testing), BatchEnsemble and Masksembles, were implemented and tested on two standard data collections: MSLRWEB10K and YAHOO!. The results of the experiments indicated that some of these ensemble approaches, specifically Masksembles and BatchEnsemble, outperformed the original AllRank in metrics such as NDCG@1, NDCG@5 and NDCG@10, although they were more costly in terms of training and testing time. In conclusion, the research reveals that the application of sub-networks ensemble in L2R models is a promising strategy, especially in scenarios where latency time is not critical. Thus, this work not only advances the state of the art in L2R, but also opens up new possibilities for improvements in effectiveness and efficiency, inspiring future research into the use of sub-networks ensemble in L2R.