Correlation-based Oversampling aided Cost Sensitive Ensemble learning technique for Treatment of Class Imbalance

Abstract

ABSTRACT The issue of class imbalance and its consequences over the conventional learning models is a well-investigated topic, as it highly influences performances of real-life classification tasks. Amongst the available solutions, Synthetic Minority Oversampling Technique (SMOTE) imprints efficacy in balancing the data through synthetic minority instance generation. However, SMOTE suffers from the drawback of redundant data generation owing to uniform oversampling rate in regard to which, SMOTE with a customised oversampling rate has been investigated recently. In parallel to this, ensemble learning approaches are quite effective in improving prediction abilities of a set of weak classifiers through adaptive-weighted training. However, it does not account the imbalanced nature of the data during training. Through this paper, Correlation-based Oversampling aided Cost Sensitive Ensemble learning (CorrOV-CSEn) is proposed by integrating correlation-based oversampling with the AdaBoost ensemble learning model. The correlation-based oversampling entails to define a customised oversampling rate and a suitable oversampling zone while a misclassification ratio-based cost-function is introduced in the AdaBoost model to administer adaptive learning of imbalanced cases. CorrOV-CSEn is evaluated against 13 state-of-the-art methods by using 8 simulation datasets. The experimental results establish CorrOV-CSEn to be effective than the state-of-the-art methods in resolving the concerned issues.