Amostragem aleatória e extensões para predição de eventos raros

Abstract

In classification problems, the prediction of rare events, that is, when the class of interest is underrepresented„ is often a difficult issue to solve. Classical versions of algorithms suffer several problems when being trained when the response variable is unbalanced, and certain metrics, such as accuracy, lose value when comparing different models. In this dissertation, we present different random sampling techniques and their applications in extensions of ensemble techniques that aim to solve this dilemma. Although extensions exist for most methods used in multi-class problems, we focus on their use for dichotomous problems. In addition, we performed simulations on databases seeking to observe advantages and shortcomings of the methods used, with emphasis on a credit concession database, where the imbalance is severe (below 5%)