Método de autogeração de protótipos para tarefas de classificação de dados baseado em entropia da informação

• Autor(a) principal: Manasrarla, Alberto
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações do Mackenzie
• Texto Completo: http://dspace.mackenzie.br/handle/10899/26584
• topic: seleção de protótipos; redução de dados; classificação de dados; entropia de informação; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO::BANCO DE DADOS

Currently, data generated in large volume, different contexts of application and format have significantly impacted on different types of machine learning algorithms, especially those of predictive modeling. Systems that involve data with these characteristics cannot afford the requirements of time, storage, and real-time processing. Under these conditions, classifying, understanding, or compacting the available data can become an expensive task. The problem is particularly acute when using distance-based learning algorithms , such as the Nearest-Neighbor or NN rule. The basic NN algorithm in a data classification task is to measure the distance between an instance under analysis, and all available instances which denominates the training set. It means that instances in the training set should always remain stored for future analysis, even those that may contain noise or may not be representative for analysis. Given this circumstance, the prototype selection is presented as a proposal for data reduction and consequently minimizes the problems of computational complexity, storage requirements, and noisy data in datasets. The prototype selection aims to find a representative dataset with a smaller dimension compared to the original one and can eventually use data generation in situations of unbalanced classes. However, most of these methods can perform anexhaustive process to find the reduced dataset because they do not have precise control over the prototype generation process. Then,a lack of control in the prototype generation process can lead to producing a reduced non-optimal dataset size. In the context of this, the present work proposes a new method for self-generating prototypes, in which it makes use of ideas from information theory, precisely the concept of information entropy, to determine the optimal number of prototypes to form the reduced dataset. In the data classification task with the k-Nearest Neighbors or kNN algorithm, the proposed method makes it possible to accelerate the time of classification without significantly reducing the quality of classification. In a comparative study with another self-generating prototype technique of the literature,SSGA (Steady-State Genetic Algorithm) for prototype selection, under quantitative and qualitative analysis by using different datasets (public and artificial),the results for the proposed method are higher on average 85.71 percent for the classification time with an average rate of data reduction of 77 percent, and an increase in accuracy efficiency of 3 percent for the kNN classifier.