Unified Classifier Stacking: A Comprehensive Ensemble Approach to Cyberbullying Detection

Cyberbullying is a growing problem in digital spaces, making the development of accurate detection strategies more critical than ever. Despite various models created in recent years, accurately predicting such harmful behavior remains a challenge, highlighting the need for more advanced solutions. In this study, we introduce a stacked ensemble approach to improve cyberbullying detection. Our model combines Decision Tree (DT), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN) as base learners, with Logistic Regression (LR) acting as the meta-learner. Using a dataset of Twitter posts, we performed crucial preprocessing steps like lowercasing, removing stopwords, and stemming. We then applied TF-IDF vectorization, Random Over Sampling (ROS), and K-Fold cross-validation to prepare and evaluate the model. The stacked ensemble approach demonstrated notable improvements, achieving 91.43% accuracy, 91.99% precision, 90.46% recall, a 91.22% F1-score, and an AUC of 96.77%. These results confirm that combining diverse algorithms provides a robust method for detecting cyberbullying and could significantly contribute to making online environments safer.