Hybrid BERT-XGBoost Framework for Early Detection and Classification of Online Cyberbullying across Social Media

Abstract

Cyberbullying via social media is a constant digital safety issue because the content can be widely shared and openly visible and can have a negative impact on users before it is removed by manual moderation. Current detection models are mostly based on shallow lexical features or transformer-only classifiers, resulting in low-level accuracy and explainability. This study introduces a Hybrid BERT–XGBoost model to detect cyberbullying in short social media texts, which combines the strengths of both models. The contextual sentence embeddings are extracted using BERT and the auxiliary linguistic and behavioral features are extracted in parallel, such as sentiment polarity, profanity score, punctuation intensity, capitalization ratio, hashtag usage, mention count, emoji frequency, and post length. XGBoost is used for the classification of the fused representation. The model is tested on stratified training, validation, and testing splits, compared to a baseline model, ablated, tested with macro-F1, weighted-F1, ROC-AUC, early detection recall, and grouped explainability. The proposed framework achieved 96.18% accuracy, 96.05% macro-F1, 96.16% weighted-F1, 95.88% early detection recall, and 98.42% macro-AUC. It performs better than the BERT + Dense baseline, which obtained 94.31% accuracy and 94.08% macro-F1 score, demonstrating the advantage of fusion of contextual and auxiliary features. The framework provides an interpretable, practical and category-aware solution for early detection of cyberbullying, but further research is needed to validate the framework in multiple languages, modalities and in conversations.