A novel rule-based oversampling approach for imbalanced data classification

Abstract

When confronted with imbalanced datasets, traditional classifiers frequently struggle to correctly categorize samples from the minority class, adversely impacting the overall predictive performance of machine learning models. Current oversampling techniques generally focus on data interpolation through neighbor selection, often neglecting to uncover underlying data structures and relationships. This study introduces a novel application for RuLer, an algorithm originally developed for identifying sound patterns in the artistic domain of live coding. When adapted for data oversampling (as Ad-RuLer), the algorithm shows significant promise in addressing the challenges associated with imbalanced class distribution. We undertake a thorough comparative evaluation of Ad-RuLer against established oversampling algorithms such as SMOTE, ADASYN, Tomek-links, Borderline-SMOTE, and KmeansSMOTE. The evaluation employs various classifiers including logistic regression, random forest, and XGBoost, and is conducted over six real-world biomedical datasets with varying degrees of imbalance.