Classifying imbalanced data using SMOTE based class-specific kernelized ELM

Abstract

In machine learning, a problem is imbalanced when the class distributions are highly skewed. Imbalanced classification problems occur usually in many application domains and pose a hindrance to the conventional learning algorithms. Several approaches have been proposed to handle the imbalanced learning. For example, Weighted kernel-based SMOTE (WKSMOTE) and SMOTE based class-specific extreme learning machine (SMOTE-CSELM) are recently proposed algorithms that use the minority oversampling to handle imbalanced learning. It has been illustrated in Raghuwanshi and Shukla (Knowl-Based Syst 187(104):814, 2020) that our recently proposed classifier, SMOTE-CSELM outperforms the other state of art classifiers for class imbalance learning. One drawback of SMOTE-CSELM is the performance fluctuation due to the random initialization of weights between the input and the hidden layer. To handle this problem, this work proposes SMOTE based class-specific kernelized extreme learning machine (SMOTE-CSKELM), which uses the Gaussian kernel function to map the input data to the feature space. The proposed work has the advantage of both the minority oversampling and the class-specific regularization coefficients. SMOTE-CSKELM with the Gaussian kernel function also handles the non-optimal hidden node problem associated with the sigmoid node based variants of ELM. To increase the significance of the specific region corresponding to the minority class in the decision boundary, the synthetic minority oversampling technique (SMOTE) is applied to generate synthetic instances for the minority class to balance the training dataset. The proposed work has comparable training time in contrast with the kernelized weighted extreme learning machine (KWELM) for imbalanced learning. The proposed method is determined by employing benchmark real-world imbalanced datasets. The extensive experimental results report that the proposed method outperforms compared to the other state-of-the-art methods for imbalanced learning.