Detecção de mudança de conceito baseada em aprendizado ativo

Abstract

Current machine learning techniques used for prediction tasks rely on the assumption that the environment where data is generated is static and supervised. However, most of the real-world problems present dynamic and semi-supervised environments, which invalidate this assumed assumption. In these environments, it is possible the occurrence of a phenomenon known in the literature as concept drift. The main characteristic of such a phenomenon is that the relationship between instances’ features and their true classes changes over time. The primary effect of the concept drift occurrence is the decrease on classifier performance, which leads the classifier to be obsolete for the task. There are methods in the literature that deal with concept drift implicitly or explicitly. The main drawback to blind (implicit) methods is the cost on retraining the classifier constantly, even with no concept drift occurrence, while the informed (explicit) methods generally work based on classifier performance decrease, or based on assumptions about the samples distribution. In order to try to overcome these drawbacks, in this work we propose to use density variation of the most significant instances as an explicit trigger for concept drift detection. Density variation measurement is based on Active Learning, and is calculated from virtual margins projected onto the input space according to the classifier confidence. Focusing on demonstrating the validity of the proposed method, called Concept Drift Detection Method Based on Active Learning (DMAA), were have carried out experiments divided into two series. In the first series, DMAA was investigated on six databases, which represent four synthetic and two real problems. The obtained results show that the proposed method achieved 95.45% of drift detection rate on synthetic databases, and 16.5% as mean error rate in both synthetic and real databases. In the second series of experiments, DMAA was compared to three baselines, including two supervised drift detectors and one Active Learning-based method. The obtained results show that DMAA reached statistically significant better recognition rates in the majority of databases. Besides, it reduced the amount of labeled instances needed to keep the system updated.