Estratégias evolucionárias para otimização no tratamento de dados ausentes por imputação múltipla de dados

Abstract

The data analysis process includes information acquisition and organization in order to obtain knowledge from them, bringing scientific advances in various fields, as well as providing competitive advantages to corporations. In this context, an ubiquitous problem in the area deserves attention, the missing data, since most of the data analysis techniques can not deal satisfactorily with this problem, which negatively impacts the final results. In order to avoid the harmful effects of missing data, several studies have been proposed in the areas of statistical analysis and machine learning, especially the study of Multiple Data Imputation, which consists in the missing data substitution by plausible values. This methodology can be seen as a combinatorial optimization problem, where the goal is to find candidate values to substitute the missing ones in order to reduce the bias imposed by this issue. Metaheuristics, in particular, methods based in evolutionary computing have been successfully applied in combinatorial optimization problems. Despite the recent advances in this area, it is perceived some shortcomings in the modeling of imputation methods based on evolutionary computing. Aiming to fill these gaps in the literature, this thesis presents a description of multiple data imputation as a combinatorial optimization problem and proposes imputation methods based on evolutionary computing. In addition, due to the limitations found in the methods presented in the recent literature, and the necessity of adoption of different evaluation measures to assess the imputation methods performance, a multi-objective genetic algorithm for data imputation in pattern classification context is also proposed. This method proves to be flexible regarding to data types and avoid the complete-case analysis. Because the flexibility of the proposed approach, it is also possible to use it in other scenarios such as the unsupervised learning, multi-label classification and time series analysis.