Considerando o ruído no aprendizado de modelos preditivos robustos para a filtragem colaborativa

• Ano de defesa: 2018
• Autor(a) principal: Carmo, Filipe Braida do
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio de Janeiro; Brasil; Instituto Alberto Luiz Coimbra de Pós-Graduação e Pesquisa de Engenharia; Programa de Pós-Graduação em Engenharia de Sistemas e Computação; UFRJ
• Palavras-chave em Português: Filtragem colaborativa; Ruído de classe; CNPQ::ENGENHARIAS
• Link de acesso: http://hdl.handle.net/11422/12978

In Recommendation Systems, it is named natural noise the inconsistencies that are introduced by a user. These inconsistencies affect the overall performance. Until then, data cleansing proposals have emerged with the objective to identify and correct these inconsistencies. However. approaches that consider noise in the learning process present a superior quality. Meanwhile, procedures for changing the cost function have arisen whose solution for the minimization of this with noisy data corresponds to the same solution using the original function with noiseless data. However, these procedures are dependent on previews knowledge of the noise distribution and in order to estimate it, certain assumptions regarding data are required. These conditions are not satisfied in collaborative filtering. In this work it is proposed to use these cost functions to construct a predictive model that considers noise in its learning. In addition, we present: (a) a class noise generation heuristic for collaborative filtering problems; (b) a baseline noise quantitative analysis; (c) robustness analysis of predictive models. In order to validate the proposal, three most representative datasets were selected for the problem. For such datasets, comparisons were made with state-of-the-art. Our results indicate that the proposal obtains superior prediction quality to the other methods in all the datasets and maintains a competitive robustness even when compared with the model that knows a priori the generator of the noise. Finally, a new direction is opened for methods that consider noise to the learning process of predictive models for collaborative filtering.