Aprendizagem semi-supervisionada aplicada à engenharia financeira

Abstract

Semi-supervised learning had become, recently, a good alternative toimprove generalization capacity in machine learning models. The approach is generally used in problems that labeled samples are hard tobe obtained and unlabeled, in turn, are plenty and easily collected; the semisupervised training algorithm tries to generate the separation surface between the two classes based also on unlabeled data. A classic example is the web sites classification: give a trustworthy label to a sample is time consuming. Nonetheless, a crawler can quickly collect a great number of unlabeled samples. This work shows a different application of semi-supervised learning, because in this case, it is possessed a long labeled training set (financial time series can be easily downloaded from internet), however, the quality of this set is put under prove: the labels are trustworthy? Through the semi-supervisedapproach it was possible minimize the noise data from training set, improving the results obtained.