A hybrid framework combining large language models and Shapley-based methods for explainable cyberbullying detection

Cyberbullying remains a pervasive threat across social media platforms, highlighting the need for reliable detection systems and transparent decision-making processes. This work presents a comprehensive framework for cyberbullying detection that integrates two complementary eXplainable Artificial Intelligence (XAI) techniques based on Large Language Model (LLM) explanations and a SHAP value–based interpretability. The framework leverages different models for message classification (including Decision Tree, Gradient Boosting and Multi-Layer Perceptron) to identify harmful content analyzing message-level, user-level, linguistic and stylistic features. The proposed architecture is designed to pair predictive performance with explainability, offering both natural-language justifications and feature-attribution visualizations. To assess the usability and perceived quality of the explanations, a structured evaluation has been conducted with 100 participants, including students, IT professionals, and researchers. Responses have been analyzed with respect to the Co-12 explainability properties, enabling an in-depth comparison between the two XAI approaches. Results indicate that SHAP provides more detailed content-based insights, while LLM-based explanations offer clear and user-oriented presentation. The findings demonstrate the importance of combining complementary XAI techniques to support transparency in current social networks.