BES-Optimized SMOTE Variant to Improve Dataset Scaling for Enhanced Privacy-Preserving Classification

The research paper undertakes a comprehensive comparative assessment of various Synthetic Minority Over-sampling Technique (SMOTE) variants in the context of medical datasets. The study is dedicated to mitigating the challenges posed by imbalanced data in medical applications. The performance of SMOTE, SMOTE-ENN, Borderline SMOTE, SMOTE-Tomek Links, Safe Level SMOTE, Density-based SMOTE, SMOTE-Rough Set Theory, and a developed BES-Optimized SMOTE are evaluated across five distinct medical datasets such as heart disease, arrhythmia, sick euthyroid, hypothyroid and breast cancer. Additionally, the K-Nearest Neighbors (KNN), Decision Forest (DF), Random forest (RF), and Support Vector Machine (SVM) algorithms are utilized to analyze the performance of SMOTE variants on each dataset, where the developed Bald Eagle search (BES) optimized SMOTE shows superior performance than other SMOTE variants, indicating its effectiveness in generating balanced scaled datasets. Subsequently, the BES-optimized SMOTE is chosen as the most effective SMOTE variant to scale and balance the respective datasets. Following a thorough examination of results, datasets that are appropriately scaled with the selected SMOTE variant are employed for ensuing research endeavors. These refined and balanced datasets will serve as the foundation for performance analysis of our future work on efficient privacy preserving KNN classification. To facilitate this, the scaled datasets in this work will be encrypted using the Paillier Encryption Scheme.