Hybrid Models with Dynamic Fusion for Cyberbullying Detection: A Comparative Analysis

Cyberbullying is one of the well-known problems, especially with modern social media systems, as more people use this form of connectivity. It often takes the route of harassment and threatening or humiliations that can bring emotional and psychological harm to targeted victims. Efficient ways of detecting harmful content with rapid and accurate classifications are necessary solutions for this problem. This research aims to boost cyberbullying detection using a dynamic fusion approach to aggregate multiple machine learning models. Here, we combine Hybrid models of Random Forest (RF) and Naive Bayes (NB), LightGBM and Naive Bayes, Naive Bayes and Gradient Boosting, Logistic Regression (LR) and Random Forest for exploiting the respective strengths of many types of classification algorithms. These hybrid models were trained over a dataset of social media, which was balanced and labeled into four categories of cyberbullying. The fusion approach dynamically weighs model outputs and integrates them for more accurate classification. Our findings reveal that the Naive Bayes and LightGBM hybrid model achieves the highest accuracy outperforming other hybrid models. These indicate that hybrid models have the capability to effectively identify multi-class cyberbullying on digital channels.