Farthest SMOTE: A Modified SMOTE Approach

Abstract

Class imbalance problem comprises of uneven distribution of data/instances in classes which poses a challenge in the performance of classification models. Traditional classification algorithms produce high accuracy rate for majority classes and less accuracy rate for minority classes. Study of such problem is called class imbalance learning. Various methods are used in imbalance learning applications, which modify the distribution of the original dataset by some mechanisms in order to obtain a relatively balanced dataset. Most of the techniques like SMOTE and ADASYN proposed in the literature use oversampling approach to handle class imbalance learning. This paper presents a modified SMOTE approach, i.e., Farthest SMOTE to solve the imbalance problem. FSMOTE approach generates synthetic samples along the line joining the minority samples and its ‘k’ minority class farthest neighbors. Further, in this paper, FSMOTE approach is evaluated on seven real-world datasets.