A Novel Method for Imbalanced Data Classification based on Label Reassignment

Imbalanced data classification is one of the most challenging supervised learning problems in machine learning. Several data-sampling methods have been proposed in the last decade employing over-sampling and under-sampling techniques. Over-sampling technique engenders artificial minority-class instances that creates overfitting problems. However, under-sampling techniques suffer from informative information lost by removing major-part of majority-class instances. In this paper, we have proposed a novel approach for imbalanced data classification using label reassignment of misclassified majority-class instances. The idea is to consider the misclassified majority-class instances as minority-class instances to make the dataset balanced. We conceptualise the reassignment of majority-class instances in overlapping region to minority-class instances in Label Reassignment. We have conducted experiment on 11 imbalanced datasets and compared the performance of proposed method with well-known Synthetic Minority Oversampling Technique (SMOTE) method. The proposed method outperforms SMOTE in most of the experimented datasets with a significant improvement in terms of Recall and F1 Score.