ABSTRACT Class imbalance is a significant problem in classification tasks for various data types, such as tabular data and images. Data imbalance can lead to biased models that fail to correctly identify minor classes. This paper presents the results of experimental studies of weighting methods for classical ensemble algorithms applied to tabular data, as well as for neural networks used for image classification. The results demonstrate that the inverse frequency method, despite being widely adopted, does not consistently yield optimal classification performance. General recommendations for addressing class imbalance in both tabular and image data are provided. The findings suggest that alternative weighting strategies beyond standard implementations warrant exploration and should be integrated into hyperparameter optimisation.
Lukashevich et al. (Thu,) studied this question.