A Preprocessing Approach for Class-Imbalanced Data Using SMOTE and Belief Function Theory

Overview & Synopsis

Dealing with imbalanced datasets at the preprocessing level is an efficient strategy used by many methods to re-balance the data and improve classification performance. Specifically, SMOTE is a popular oversampling technique which modifies the training data by adding artificial minority samples. However, SMOTE may create instances in noisy and overlapping areas, far from safe regions. To tackle this issue, we propose SMOTE-BFT, in which we use the belief function theory to remove generated minority instances that are not in safe regions. After applying SMOTE, each generated minority instance is represented by an evidential membership structure, which provides detailed information about class memberships. Rules based on the belief function theory are then enforced to detect and remove generated instances that are in noisy and overlapping regions. Experiments on noisy artificial datasets show that our proposal significantly outperforms other popular oversampling methods.