FS-SMOTE: An Improved SMOTE Method Based on Feature Space Scoring Mechanism for Solving Class-Imbalanced Problems
Yongjie Huang · IEEE Access · 2025
Class-imbalance problems have become a key challenge in machine learning, often results in training too many majority samples and learning too few minority samples. The Synthetic Minority Oversampling Technique (SMOTE) has been proven to be one of the most effective methods for dealing with class-imbalanced problems, specifically by synthesizing new minority samples to achieve approximately equal numbers of majority and minority samples. However, traditional SMOTE and some of its improved methods have their own limitations, such as the risk of synthesizing noisy samples and the problems caused by ignoring data distribution. To overcome these shortcomings, we propose an improved SMOTE algorithm based on feature space scoring mechanism (FS-SMOTE). FS-SMOTE establishes a scoring mechanism based on feature space and select minority samples based on probability accounting to the scores to participate in the subsequent synthesis process. Experiments on several benchmark datasets have shown that the FS-SMOTE method is significantly superior to the traditional SMOTE and three improved SMOTE methods in terms of evaluation metrics.