Class imbalance is one of the most challenging problems for machine learning algorithms. The most commonly used classification algorithms also suffer from the curse of highly imbalanced data. Numerous methods have been proposed in the literature to address the class imbalance problem; however, many of these approaches are complex and often introduce unnecessary noise into the data. Moreover, effective oversampling techniques that can handle datasets containing both nominal and continuous features remain limited. Although SMOTE-ENC (SMOTE for Encoded Nominal and Continuous features) can capture the relationship between categorical features and the target variable, it does not explicitly enforce the decision boundary. To address these challenges, we propose a new oversampling approach, SVMSMOTE-ENC, for nominal and continuous features, based on a combination of a Support Vector Machine (SVM) classifier and SMOTE-ENC. Our method modifies the encoding mechanism of SMOTE-ENC and replaces its random selection of observations with a targeted selection of more informative minority class instances. An SVM classifier is trained to find the borderline between the minority and the majority classes and synthetic minority samples are generated around the SVM decision boundary. Our approach was evaluated on six standard imbalanced datasets, and the performance of SVMSMOTE-ENC was compared with SMOTE-NC and two state-of-the-art oversampling methods. The experimental results show that our proposed method provides better performance in terms of F1-score and balanced accuracy.
Sam Mohsen Fallah (Mon,) studied this question.