Empowering One-vs-One Decomposition with Ensemble Learning for Multi-Class Imbalanced Data
MetadataShow full item record
AuthorZhang, Zhongliang; Krawczyk, Bartosz; García López, Salvador; Rosales-Pérez, Alejandro; Herrera Triguero, Francisco
Multi-class classificationImbalanced dataEnsemble learningBinary decompositionClassifier combination
Published version: Zhang, Z., Krawczyk, B., Garcìa, S., Rosales-Pérez, A., & Herrera, F. (2016). Empowering one-vs-one decomposition with ensemble learning for multi-class imbalanced data. Knowledge-Based Systems, 106, 251-263. [https://doi.org/10.1016/j.knosys.2016.05.048]
SponsorshipNational Natural Science Foundation of China (NSFC) 61273204; CSC Scholarship Program (CSC) 201406080059; Polish National Science Center UMO-2015/19/B/ST6/01597; Spanish Government TIN2014-57251-P; Andalusian Research Plan P10-TIC-6858 P11-TIC-7765; Consejo Nacional de Ciencia y Tecnologia (CONACyT) 329013
Multi-class imbalance classification problems occur in many real-world applications, which suffer from the quite different distribution of classes. Decomposition strategies are well-known techniques to address the classification problems involving multiple classes. Among them binary approaches using one-vs-one and one-vs-all has gained a significant attention from the research community. They allow to divide multi-class problems into several easier-to-solve two-class sub-problems. In this study we develop an exhaustive empirical analysis to explore the possibility of empowering the one-vs-one scheme for multi-class imbalance classification problems with applying binary ensemble learning approaches. We examine several state-of-the-art ensemble learning methods proposed for addressing the imbalance problems to solve the pairwise tasks derived from the multi-class data set. Then the aggregation strategy is employed to combine the binary ensemble outputs to reconstruct the original multi-class task. We present a detailed experimental study of the proposed approach, supported by the statistical analysis. The results indicate the high effectiveness of ensemble learning with one-vs-one scheme in dealing with the multi-class imbalance classification problems.