A Hybrid Preprocessing Approach for the Classification of Class Imbalanced Data

Models trained with class imbalanced data become biased towards the majority class data and as a consequence it results in poor performance while predicting the minority class data. So this situation should be handled carefully. In data level preprocessing, several methods have been introduced, under-sampling and over-sampling are the fundamental of them. The problem with under-sampling is the significant loss of information from majority class data and the problem with over-sampling is the overfitting of minority class data. In this work, we have proposed a hybrid approach, namely HPACID (Hybrid Preprocessing Approach for Class Imbalance Data), combined of over-sampling and ensemble approach, to minimize the mentioned cons. In HPACID, boundary values, where most misclassification occurs, are handled in a unique way. Later ensemble approach is used that reduces loss of information from the majority class as much as possible. We have used Decision Tree (DT) classifier to compare the performance of HPACID with four state-of-the-art methods, such as random over sampling (ROS), random under sampling (RUS), synthetic minority over sampling technique (SMOTE) and EHSEL method. We have used five-fold cross validation along with F1-score, area under the ROC curve (AUC), and G-mean for validating the results. For ten class imbalanced dataset, our proposed method consistently outperforms other state-of-the-art methods in terms of F1-score, AUC, and G-mean, achieving the highest scores in 70%, 100%, and 100% of the time, respectively.