Um framework semissupervisionado para classificação de dados em fluxos contínuos

Abstract

Data stream applications receive a large volume of data quickly, and they need to process them sequentially. In these applications, the data may change during the use of the model; in addition, the number of instances whose label is known may not be sufficient to generate an effective model. Semi-supervised learning can be used to suppress the difficulty of the small number of instances labelled. Also, an ensemble of classifiers can assist in the concept drift detection. So, in this work, we proposed a framework to perform the semi-supervised classification in tasks in a data stream context, using an approach based on an ensemble of classifiers. This framework use an ensemble to evaluate itself and determine when a new classifier must be trained to update the pool, during the classification process. In order to evaluate the effectiveness of this proposal, empirical tests are carried out with eleven databases using two different batches sizes, nine supervised approaches (three simple classifiers and six ensembles), using the metrics accuracy, precision, recall and F-Score. When assessing the number of instances processed, the supervised approaches achieved practically stable performance, while the proposal showed an improvement of 8.28% and 3.81% using 5% and 10% of labelled instances, respectively. Finally, the results of this research are promising and the proposed framework achieve results equal or better in 118 out of 198 (60%).