An efficiency curve for evaluating imbalanced classifiers considering intrinsic data characteristics: Experimental analysis

Balancing the accuracy rates of majority and minority classes is challenging in imbalanced classification. Furthermore, data characteristics have a significant impact on performance classifiers, which are generally neglected by existing evaluation methods. The objective this study to introduce new criterion comprehensively evaluate classifiers. Specifically, we an efficiency curve that established using envelopment analysis without explicit inputs (DEA-WEI), determine trade-off between benefits improved class cost reduced accuracy. In sequence, analyze ratio typical Empirical analyses 68 reveal traditional classifiers such as C4.5 k-nearest neighbor more effective disjunct data, whereas ensemble undersampling techniques for overlapping noisy data. cost-sensitive decreases dramatically when increases. Finally, investigate reasons different efficiencies recommend steps select appropriate based characteristics.