A Comparative Study of Machine Learning and Transfer Learning Approaches for Cyber Bullying Detection in Social Networks

The pervasive issue of cyberbullying on social media sites has a fresh answer, thanks to this study. Our research presents a powerful automated approach for cyberbullying detection utilizing ICTs, including Conventional Machine Learning (CML) and Transfer Learning (TL). Our methodology incorporates a wide range of features using the AMiCA dataset with structured annotations. These characteristics include textual, sentimental, and emotional indicators, as well as contextual and static word embeddings, psycholinguistics, term lists, and toxicity markers. Specifically, when compared to alternative solutions in terms of F-measure, DistilBert embeddings perform exceptionally well. Logistic Regression outperforms LinearSVC in the context of CML because of how well it deals with high-dimensional features. Once again, DistilBert exhibited the highest F-measure when it came to Transfer Learning, which makes use of fine-tuned pre-trained language models.