Classifying multiclass imbalanced data using generalized class-specific extreme learning machine

Learning from the imbalanced problem is among the most attractive issues in the contemporary machine learning community. However, the extensive majority of attention in this domain is given to the two-class imbalanced problems, while their much more complex multiclass counterparts are comparatively unexplored. It has been shown (Huang et al. in IEEE Trans Syst Man Cybern B (Cybern) 42(2):513–529, 2012) that extreme learning machine (ELM) achieves much better generalization performance compared to support vector machine (SVM) and least-squares support vector machine (LS-SVM) for multiclass classification problems. On this account, this work proposes a novel generalized class-specific extreme learning machine (GCS-ELM), the extension of our recently proposed, class-specific extreme learning machine (CS-ELM) to address the multiclass imbalanced problems more effectively. The proposed GCS-ELM can be applied directly to the multiclass imbalance problems. The proposed method also has reduced computational cost compared to the weighted extreme learning machine (WELM) for multiclass imbalance problems. The proposed method uses class-specific regularization coefficients, which are computed by employing class distribution. The proposed method has lower computational overhead compared to the class-specific cost regulation extreme learning machine (CCR-ELM). The proposed work is assessed by using benchmark real-world imbalanced datasets downloaded from the well-known KEEL dataset repository and synthetic datasets. The experimental results, supported by the extensive statistical analysis, demonstrate that GCS-ELM is capable to improve the generalization performance for multiclass imbalanced classification problems.