Evolutionary Cost-Sensitive Balancing: A Generic Method for Imbalanced Classification Problems

Efficient classification under imbalanced class distributions is currently of interest in data mining research, considering that traditional learning methods often fail to achieve satisfying results in such domains. Also, the correct choice of the metric is essential for the recognition effort. This paper presents a new general methodology for improving the performance of classifiers in imbalanced problems. The method, Evolutionary Cost-Sensitive Balancing (ECSB), is a meta-approach, which can be employed with any error-reduction classifier. It utilizes genetic search and cost-sensitive mechanisms to boost the performance of the base classifier. We present evaluations on benchmark data, comparing the results obtained by ECSB with those of similar recent methods in the literature: SMOTE and EUS. We found that ECSB boosts the performance of traditional classifiers in imbalanced problems, achieving ~45% relative improvement in true positive rate (\(\text {TP}_{\text {rate}}\)) and around 16% in F-measure (FM) on the average; also, it performs better than sampling strategies, with ~35% relative improvement in \(\text {TP}_{\text {rate}}\) and ~12% in FM over SMOTE (on the average), similar \(text{TP}_{\text {rate}}\) and geometric mean (GM) values and slightly higher area under de curve (AUC) values than EUS (up to ~9% relative improvement).