Early Detection of Cyberbullying through Explainable Artificial Intelligence: A Lightweight Model for Intervention in Educational Environments

Abstract

Cyberbullying in educational settings, fueled by the widespread use of social media, represents a growing threat to students’ emotional and academic well-being. In response, this study proposes a lightweight and explainable artificial intelligence model for the early detection of cyberbullying in digital comments. The objective was to design an automated, efficient, and interpretable system using the DistilBERT model within an MLOps framework, ensuring traceability, scalability, and continuous integration. The methodology included data collection from Twitter, text preprocessing, stratified supervised training, and evaluation using standard classification metrics. The results demonstrate that, when trained on 100% of the dataset, the model achieved a precision of 0.87, a recall of 0.83, and an average loss of 0.235—showing significant improvements over configurations using only 20% of the data. Qualitatively, the model successfully identified offensive language patterns with varying levels of subtlety and ambiguity. The integration of SHAP for explainability enabled real-time interpretation of predictions, enhancing the model’s transparency and trustworthiness. The study concludes that the proposed approach is suitable for implementation in schools and educational platforms, offering an accessible, interpretable, and effective tool for cyberbullying prevention. Future work is encouraged to extend this framework to multilingual models and multimodal analysis for broader applicability.