Studying the perturbation-based oversampling technique for imbalanced classification problems

Abstract

This study aims to implement, analyze, and compare the effectiveness of a novel technique known as Perturbation-Based Oversampling (POS). This technique is designed to address class imbalance in machine learning by augmenting the minority class instances by strategically perturbing features using a hyperparameter ’p’. Two additional variations, namely POS 1.0 and POS 2.0, have been proposed as extensions of the original POS approach. Detailed experiments have been conducted across diverse datasets, presenting a comprehensive performance evaluation in terms of precision when compared to a selection of established methods designed to tackle unbalanced classification challenges.