Novel visual approaches for attribute analysis, selection, and prediction

• Ano de defesa: 2020
• Autor(a) principal: Júnior, Erasmo Artur da Silva
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Análise de espaço de atributos; Análise visual; Análise visual preditiva; Attribute space analysis; Data visualization; Feature selection; Predictive visual analytics; Seleção de atributos; Visual analytics; Visualização de dados
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-31082020-175620/

While data collection and storage capabilities grow widely nowadays, the general ability to process and analyze large amounts of data increases at a slower rate. This asynchrony introduces new challenges touching methods for large amounts of data, such as the ones in data mining, statistics, and machine learning. To help addressing this gap, visual approaches have been proposed to combine human capabilities with consolidated solutions in the development of interactive tools that allow a more in-depth investigation of the data. A substantial amount of visual approaches has focused on items-based techniques, where the data items represent the first-order objects. Nevertheless, valuable knowledge frequently appears from observations of relationships between attributes of these data items, such as the relationships between numerical and categorical variables, which often encode relevant information. In this context, a visual analysis approach for attribute space exploration is paramount, both when there are hypotheses of correlations that must be confirmed, and also in cases where such relationships are unknown or unforeseen. In this Thesis, we propose an approach for attribute analysis based on the simultaneous presentation of multiple correlations through a point-based visualization aiming to build cognitive maps of these relationships to the end-user. Also, the analysis process then supports additional tasks such as feature selection and the development of prediction models based on a target outcome. We show the efficiency of the approaches through a series of case studies and usage scenarios involving real data sets in distinct contexts.