Uma abordagem para imputação de valores faltantes em problemas de classificação hierárquica multirrótulo

• Ano de defesa: 2021
• Autor(a) principal: Santana, Alvaro Mateus
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Ponta Grossa; Brasil; Programa de Pós-Graduação em Ciência da Computação; UTFPR
• Palavras-chave em Português: Aprendizado do computador; Sistemas de coleta automática de dados; Classificação; Análise de regressão; Algorítmos; Machine learning; Automatic data collection systems; Classification; Regression analysis; Algorithms; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; Engenharia/Tecnologia/Gestão
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/28969

Missing data are problems commonly faced by machine learning (ML) algorithms due to several reasons, such as manual insertion failure, incorrect measurements of a given sensor, among others. Taking this into consideration, it becomes essential to use appropriate methods to impute missing data into datasets and make algorithm learning more efficient. The missing data problem is more challenging when it comes to databases with multi-label hierarchical classification with hierarchy structured by a Directed Acyclic Graph or DAG. This work is part of this scenario, where classes are arranged in a hierarchy, each instance may have more than one class. To solve this problem, a method of missing data imputation is created using three types of regression-based approach: linear, polynomial and multiple. The algorithm initially checks for correlation between the data, using regression only if this correlation exists, otherwise the average approach. observed values is adopted. The proposed method is divided into three steps: multi-label hierarchical verification, correlation calculation and model application. To perform the experiments, 7 databases of the Genetic Ontology with hierarchy structured in DAG format were used. The results showed that the use of regression presented the superior area under the precision-recall curve (AUPRC) metric in 3 of the tested databases when comparing the non-imputation approaches of missing data and mean of observed values. In addition, the Friedman and Wilcoxon statistical tests were performed in order to compare the results of all algorithms. The tests show a certain difference between the results, but they showed that statistically the difference is not significant.