Selecting Classifiers and Resampling Techniques for Imbalanced Datasets: A New Perspective
Khalifa Afane, Yijun Zhao · Procedia Computer Science · 2024
Developing effective binary classifiers for imbalanced datasets poses significant challenges, which are further compounded in the case of multi-class imbalanced datasets. To address this issue, numerous resampling techniques and hybrid algorithms have been introduced in the literature, aiming to enhance accuracy, precision, and other key metrics. In this paper, we conduct a series of comprehensive experiments using 38 imbalanced datasets, nine classifiers, and seven resampling techniques. We then approach this challenge from a new perspective by focusing on the performance of classifiers and resampling techniques based on the type of features present in the data: continuous or categorical. Our study aims to provide researchers and practitioners with a systematic framework for selecting an appropriate combination of resampling techniques and classifiers tailored to the characteristics of the imbalanced dataset at hand.