AN EFFECTIVE FILTERING EXTENSION TECHNIQUE OF SMOTE FOR CONTROLLED SYNTHETIC DATA GENERATION

Imbalanced dataset poses a significant classification challenge within the realm of machine learning and has gained growing prominence due to the need to handle real-world data that is usually imbalanced and skewed. Numerous resample techniques have been proposed in the literature to improve the performance by solving the imbalanced problem. But this situation becomes more complex when one class contains a significantly larger number of examples compared to the other classes. Under such circumstances, machine learning-based algorithms can accurately identify examples from the majority class but often struggle or likely fail to recognize instances from the minority class. However, these minority class examples often contain crucial and valuable information. In addition, generating new instances to balance the minority classes by curating some rules by the domain expert as in the standard SMOTE, may not be suitable for some instances leading to misclassification by the model. Therefore, this paper proposed a novel technique namely filter extension of the SMOTE algorithm based on optimised kernel trick on SVM to control the generation of balanced synthetic data in overlapped samples. The main objective of this paper is to predict the minority class instances accurately and robustly addressing the imbalance and overlapping issues. The proposed technique is validated by a rigorous testing framework utilizing a 10-fold cross-validation method to ensure a comprehensive evaluation of support vector machine (SVM) classifier. Several parameters, such as, AUC and G-mean metrics were used to ensure the accuracy, robustness and effectiveness of the proposed technique comparing to other traditional and machine learing-based methods. We experimented with highly imbalanced dataset from KEEL repository. The proposed approach outperformed the standard SMOTE and RC-SMOTE, proven the effectiveness of the proposed approach of filtering imbalance data in improving the generalizability and performance of machine learning classifiers.