Agregação de classificadores e agrupadores em problemas com classes desbalanceadas

Abstract

The combination of multiple classifiers to generate a single classifier has proven to be very useful in several application areas, as it tends to increase the accuracy of the classification process. Similarly, various studies have shown that the aggregation of clusterers can improve the quality of results compared to a single clustering solution. Furthermore, the use of clustering techniques in supervised learning tasks can provide complementary constraints that help improve the generalization ability of classifiers. In this context, the aggregation of classifiers and clusterers tends to enhance classification results, as it aims to reconcile the benefits of using both labeled and unlabeled data in the learning process. In this monograph, we aim to study three algorithms that combine clusterers and classifiers: CBCA (Classification by Cluster Analysis), DTC (Decision Templates Combiner) and CCCE (Classifier Combination using Clustering Ensemble). We apply these algorithms to three real-world datasets in which the output variables present imbalanced classes. We observe that these algorithms face difficulties in classifying instances belonging to the minority class. To address this challenge, we propose modifications to the CBCA algorithm and to the algorithm proposed by Wang et al. (2015). The analyses showed that the proposed modifications resulted in more efficient algorithms, competitive with other methods reported in the literature for handling class imbalance.