The effect of rebalancing techniques on the classification performance in cyberbullying datasets

Cyberbullying detection systems rely increasingly on machine learning techniques. However, class imbalance in cyberbullying datasets, where the percentage of normal labeled classes is higher than that of abnormal labeled ones, presents a significant challenge for classification algorithms. This issue is particularly problematic in two-class datasets, where conventional machine learning methods tend to perform poorly on minority class samples due to the influence of the majority class. To address this problem, researchers have proposed various oversampling and undersampling techniques. In this paper, we investigate the effectiveness of such techniques in addressing class imbalance in cyberbullying datasets. We conduct an experimental study that involves a preprocessing step to enhance machine learning algorithm performance. We then examine the impact of imbalanced data on classification performance for four cyberbullying datasets. To study the classification performance on balanced cyberbullying datasets, we employ four resampling techniques, namely random undersampling, random oversampling, SMOTE, and SMOTE + TOMEK. We evaluate the impact of each rebalancing technique on classification performance using eight well-known classification algorithms. Our findings demonstrate that the performance of resampling techniques depends on the dataset size, imbalance ratio, and classifier used. The conducted experiments proved that there are no techniques that will always perform better the others.