A Comparative Analysis of Resampling Techniques for Addressing Class Imbalance in Multiclass Classification
Muhammad Asif, Jawad Ali Bhatti, Muhammad Safwan, Muhammad Rafay Khan, Fahad Ahmed Siddiqui, Muhammad Maaz Rehan · 2024
This study addresses the issue of class imbalance in multiclass time series classification by using various resampling techniques. In this study, four prominent strategies are used that are Synthetic Minority Over-sampling Technique (SMOTE), Adaptive Synthetic (ADASYN) sampling, Random Under-sampling, and a hybrid method (SMOTETomek). The techniques are applied to the MIT-BIH Arrhythmia Dataset as a case study. The performance and impact of each technique are measured on classification performance. Our methodology employs a Random Forest classifier and a comprehensive set of evaluation metrics. Results demonstrate that while all resampling methods improve minority class recognition, the hybrid SMOTE-Tomek approach shows the most promising results, achieving the best balance between overall performance and class-wise metrics.