A Boosting based Adaptive Oversampling Technique for Treatment of Class Imbalance

摘要

The topic of class imbalance and its consequences have steered up the field of research for quite a long, as they bring pivotal impact over real-life scenarios such as medical disease diagnosis, fraud detection, etc. The typical solutions include data-level (undersampling or oversampling) or algorithmic-level (cost-sensitive learning) approaches. Synthetic Minority Oversampling Technique (SMOTE) has been acknowledged as one of the most effective data level solutions, but often suffers from the drawback of overfitting due to uniform oversampling rate. The ensemble learning techniques have recently emerged as effective; but can yield best results when integrated with data level solutions. In this work, a Boosting based oversampling technique is introduced with a customized oversampling rate, within an ensemble framework through cost-sensitive error formulation. The oversampling rate is tailored by using Local Covariance Matrix (LCM), while AdaBoost ensemble model with C4.5 weak learner is implemented as the ensemble framework. The work is compared with six benchmark techniques, for seven binary datasets. The experimental results prove the efficiency of the proposed work in treatment of imbalanced data.