Detecção e classificação de desvio de conceito

Abstract

Data can be used to extract useful information. Some of them are in a static environment, in which the data properties do not change over time. Others are in a dynamic environ- ment. This dynamism produces a problem known as concept drift, when the concept learned about the data changes over time. The consequence of this phenomenon is that a model trained with data before the concept change will potentially be out of date, no longer reflecting the concept and, therefore, presenting low performance. To solve this problem there are several drift detectors, but few of them can classify the drift. Therefore, this work analyzes the state of the art in concept drift classification to help algorithms better respond to drift and assist users in decision making by providing more details about what is happening with the data. To this end, a synthetic database containing different types and causes of concept drift was created. Then, a prototypical network was trained with this dataset. In order to test the developed detector, several experiments were executed, including: the comparison of the algorithm with other models, the detection and identification of drift in a synthetic set and the detection in real and synthetic datasets in classification and clustering tasks. Finally, the developed algorithm outperformed other detectors in several scenarios.