Seleção de amostras de dados menos representativas usando aprendizado ativo

Abstract

Data is increasingly available to be collected and stored, a consequence of the large amount produced by several interconnected devices. This data, for example, be used in supervised learning tasks, the purpose of which is to predict behavior based on previously labeled data. Supervised methods are used in the context of classification, such as classifying whether an email is SPAM or not, or categorizing text documents into predefined categories - sports, politics, etc. However, if the data does not have labels, on supervised learning, the possibility of use may be difficult. Labeling is a process that can be costly in a matter of time or financial resources. However, finding informative and representative examples represents a cost reduction. In this context, active learning consists of studying techniques to reduce the number of instances present in the training, selecting only the most informative ones for labeling. This work sought to explore configurations of an active learning algorithm in order to select more instances of the non-dominant class and reduce the total of selected instances. With the experiments, it was possible to increase the non-dominant class instances and reduce the quantity of dominant class instances.