Métodos de aprendizado ativo

Abstract

In the field of supervised learning, the good performance of a prediction model is generally tied to the presence of a large labelled training set. However, there are many situations where labelling an instance is expensive for financial, work and/or difficulty reasons, being prohibitive to label all observations. The use of active learning is crucial for these situations. Active learning is characterized by, through different methods, selecting and adding more informative instances to the training set of a prediction model, so that it performs well with fewer instances. In this work, we studied some active learning methods such as uncertainty sampling, query by committee, expected error reduction, variance reduction, density-weighted methods and batch-mode active learning. We also propose a new active regression method with an approach different from other methods in the literature. As part of the results, we present a simulation study to illustrate how sampling bias occurs in active learning algorithms. Finally, we explore our new active regression methodology, comparing it to other active learning methodologies and passive learning.