Multi-Level Cyberbullying Detection on Social Media Using Machine and Deep Learning Models

The rapid rise of social media has amplified the issue of cyberbullying, necessitating effective automatic detection methods. This study investigates machine learning and deep learning models—including Naive Bayes (Multinomial and Bernoulli), logistic regression, random forests, Support Vector Machines (SVM), bidirectional Long Short-Term Memory (BiLSTM) networks, convolutional neural networks (CNN), CNN-BiLSTM, and BERT—to classify tweets based on harmful content.Unlike earlier studies that primarily relied on single-level analyses, our multi-level approach provides a more robust framework for detecting cyberbullying. The first stage employs a binary classifier to distinguish cyberbullying messages from non-harmful content. In the second stage, a multi-class classifier identifies specific subtypes of cyberbullying. Leveraging a dataset of over 47,000 tweets—preprocessed through tokenization, lemmatization, and stop-word removal—the BERT model achieved a notable accuracy of 98.59% in multi-class classification. These findings highlight BERT’s potential in enhancing user safety on social media platforms.