Seleção de atributos relevantes: aplicando técnicas na base de dados do Herbário Virtual da Flora e dos Fungos

Abstract

Virtual herbariums aim to disseminate scientific information and contribute to the conservation and sustainable use of Brazilian biological resources. It currently includes 120 national herbaria and 25 herbariums from abroad, together provide more than 5,4 million records and more than one million images, in addition to several free access tools, opening space for the application of Machine Learning techniques, among them classifiers. In the Machine Learning process, Attribute Selection is part of the pre-processing of data and can correspond to 80% of the data mining phase, for this it is necessary to study the approaches used to make the selection of a subset of attributes that better generalize the basis to be induced to the model of machine learning. The objective of this work is to apply the attributes selection processes with the following filter, wrapper and embedded approaches in the National Institute of Science and Technology (NIST) - Virtual Herbarium of Flora and Fungi, this base contains 87,732 records and 51 features, with 119 collections and sub-collections, 86,967 online records, 80,513 georeferenced records, 12,073 different accepted species. The first phase of machine learning processes is the pre-processing, which will analyze the database and will result in a more general and ready basis for the application of the predictive models of classification, after the filter of the most relevant subset of attributes, the Machine Learning algorithms are applied, which in this research was: Decision Tree, Network Neural Artificial and Logistic Regression. The evaluation of the models will be through the confusion matrix using the accuracy and the analysis of the area on the ROC curve. Among the models studied, the Logistic Regression was the one that obtained the performance with a total accuracy of 77.25%, with the filter approach and 76.25% with the wrapper.