Seleção dinâmica de comitês de classificadores baseada em diversidade e acurácia para detecção de mudança de conceitos

Abstract

Many machine learning applications have to deal with classification problems in dynamic environments. This type of environment may be affected by concept drift, which may reduce the accuracy of classification systems significantly. In this context, methods using ensemble of classifiers are interesting due to the fact that ensembles of classifiers allow the design of strategies for drift detection and reaction more accurate and robust to changes. A classification system based on ensemble of classifiers may be divided into three main phases: classifier generation; single classifier or subset of classifier selection; and classifier fusion. The selection phase may be performed as a dynamic process. In this case, for each unknown sample, the individual classifier or classifier ensemble most likely to be correct is chosen to assign a label to the sample. In this work, it is proposed a method for concept drift detection and reaction based on dynamic classifier ensemble selection. The proposed method choses the expert classifier ensemble according to diversity and accuracy values. Focusing on evaluating the impact of dynamic ensemble selection guided by diversity and accuracy in terms of concept drift detection and reaction, four series of experiments were carried in this work using both synthetic and real datasets. In addition, since the proposed method is broken down into four phases: pool of ensemble classifiers generation; dynamic ensemble selection; drift detection; and drift reaction, different versions of the proposed method were investigated by varying the parameters of each phase. The results show that, in general, all these different versions attain very similar accuracy values. Besides, when compared to two baselines: (1) DDM - single classifier-based; and (2) Leveraging Bagging - classifier ensemble-based, our method outperforms both baselines since it achieved higher accuracy, lower detection delay and false detection rates, and it did not present missing detection. However, both baselines present lower time complexity. Therefore, this work shows that dynamic classifier ensemble selection guided by diversity and accuracy helps to improve detection precision and the general accuracy of classification systems employed in problems with concept drift.